merge(master): resolve conflicts for PR #703 onto current master

Preserve #702 recovery kinds (session lease shadow, stale-binding audit)
and instance_id shadow keys while integrating master #709/#720
decision-lock archive, irrecoverable provenance, and TTL-exempt
KIND_DECISION_LOCK. Keep list_states (#702) and cross-profile load APIs
(#709).

.env.example

@@ -55,3 +55,12 @@ GITEA_MCP_PROFILE=prgs
 # GITEA_MERGER_WORKTREE=/path/to/repo/branches/merge-pr456
 # GITEA_RECONCILER_WORKTREE=/path/to/repo/branches/reconcile-pr456
 # GITEA_ACTIVE_WORKTREE=/path/to/repo/branches/session-override
+
+# Durable HMAC key for irrecoverable decision-lock provenance artifacts (#709 F4).
+# REQUIRED in production native MCP processes that mint or verify recovery
+# authorization (reconciler + merger must share the same key). Hex (preferred),
+# base64, or utf-8 literal (>=16 bytes). Never generate an ephemeral per-process
+# key — cross-process / post-restart verification would fail closed.
+# GITEA_IRRECOVERABLE_AUTH_HMAC_KEY=0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef
+# Optional key version string bound into the signature (for rotation).
+# GITEA_IRRECOVERABLE_AUTH_HMAC_KEY_VERSION=v1

branch_cleanup_guard.py

@@ -7,6 +7,19 @@ from typing import Any
 
 PROTECTED_BRANCHES = frozenset({"master", "main", "dev"})
 
+# Evidence / preservation branches must never be removed by cleanup tools
+# (e.g. chore/issue-681-preserve-review-session-wip).
+_PRESERVATION_MARKERS = ("preserve", "preservation", "evidence")
+
+
+def is_preservation_or_evidence_branch(branch: str | None) -> bool:
+    """Return True when *branch* is a preservation/evidence ref that must stay."""
+    if not branch:
+        return False
+    name = str(branch).lower()
+    return any(marker in name for marker in _PRESERVATION_MARKERS)
+
+
 _RAW_BRANCH_DELETE_PATTERNS = (
     re.compile(r"\bgit(?:\s+-C\s+\S+)?\s+branch\s+-[dD]\b[^\n\r]*", re.I),
     re.compile(r"\bgit(?:\s+-C\s+\S+)?\s+push\b[^\n\r]*\s--delete\b[^\n\r]*", re.I),
@@ -72,6 +85,11 @@ def assess_merged_pr_branch_cleanup(
         reasons.append("PR head branch is missing")
     if head_branch in protected:
         reasons.append(f"branch '{head_branch}' is protected")
+    if is_preservation_or_evidence_branch(head_branch):
+        reasons.append(
+            f"branch '{head_branch}' is a preservation/evidence branch and "
+            "cannot be deleted through merged-PR cleanup"
+        )
     if head_branch in open_pr_heads:
         reasons.append("an open PR still references this head branch")
     if head_on_target is False:
@@ -96,3 +114,490 @@ def assess_merged_pr_branch_cleanup(
         "block_reasons": reasons,
         "recommended_action": "delete_remote_branch" if safe else "keep_remote_branch",
     }
+
+
+# ---------------------------------------------------------------------------
+# #687 remediation: post-delete readback + active ownership protection
+# ---------------------------------------------------------------------------
+
+READBACK_NOT_FOUND = "not_found"
+READBACK_EXISTS = "exists"
+READBACK_AUTHENTICATION = "authentication_error"
+READBACK_AUTHORIZATION = "authorization_error"
+READBACK_TRANSPORT = "transport_error"
+READBACK_UNEXPECTED = "unexpected_response"
+READBACK_AMBIGUOUS_404 = "ambiguous_not_found"
+
+# Scope of a 404: only branch-scoped absence may set verified_absent.
+NOT_FOUND_SCOPE_BRANCH = "branch"
+NOT_FOUND_SCOPE_REPOSITORY = "repository"
+NOT_FOUND_SCOPE_HOST = "host"
+NOT_FOUND_SCOPE_UNKNOWN = "unknown"
+
+ERROR_CLASS_AUTHENTICATION = "authentication"
+ERROR_CLASS_AUTHORIZATION = "authorization"
+ERROR_CLASS_TRANSPORT = "transport"
+ERROR_CLASS_UNEXPECTED = "unexpected"
+
+OWNERSHIP_CATEGORY_AUTHOR_SESSION = "author_session"
+OWNERSHIP_CATEGORY_AUTHOR_LEASE = "author_lease"
+OWNERSHIP_CATEGORY_REVIEWER_LEASE = "reviewer_lease"
+OWNERSHIP_CATEGORY_MERGER_LEASE = "merger_lease"
+OWNERSHIP_CATEGORY_CONTROLLER_LEASE = "controller_lease"
+OWNERSHIP_CATEGORY_RECONCILER_LEASE = "reconciler_lease"
+OWNERSHIP_CATEGORY_WORKTREE_BINDING = "worktree_binding"
+OWNERSHIP_CATEGORY_INVENTORY_ERROR = "ownership_inventory_error"
+
+_ROLE_TO_OWNERSHIP_CATEGORY = {
+    "author": OWNERSHIP_CATEGORY_AUTHOR_LEASE,
+    "reviewer": OWNERSHIP_CATEGORY_REVIEWER_LEASE,
+    "merger": OWNERSHIP_CATEGORY_MERGER_LEASE,
+    "controller": OWNERSHIP_CATEGORY_CONTROLLER_LEASE,
+    "reconciler": OWNERSHIP_CATEGORY_RECONCILER_LEASE,
+}
+
+_ACTIVE_OWNERSHIP_STATUSES = frozenset(
+    {"active", "live", "claimed", "in_progress", "working", "pushing", "pushed"}
+)
+_TERMINAL_OWNERSHIP_STATUSES = frozenset(
+    {"released", "abandoned", "done", "blocked", "terminal", "closed"}
+)
+_EXPIRED_STATUSES = frozenset({"expired"})
+_STALE_STATUSES = frozenset({"stale", "stale_dead_process", "stale_missing_worktree"})
+
+
+def _norm_str(value: Any) -> str:
+    return str(value or "").strip()
+
+
+def normalize_host(host: str | None) -> str:
+    """Normalize a host identity for ownership matching (no credentials)."""
+    text = _norm_str(host).lower()
+    for prefix in ("https://", "http://"):
+        if text.startswith(prefix):
+            text = text[len(prefix) :]
+    # Drop path/query if a full URL slipped through.
+    text = text.split("/", 1)[0]
+    text = text.split("?", 1)[0]
+    return text.rstrip(".")
+
+
+def ownership_category_for_role(role: str | None) -> str:
+    """Map a role kind to a non-secret ownership category label."""
+    key = _norm_str(role).lower()
+    return _ROLE_TO_OWNERSHIP_CATEGORY.get(key, f"{key or 'unknown'}_lease")
+
+
+def classify_branch_readback_http_status(
+    status_code: int | None,
+    *,
+    not_found_scope: str | None = None,
+) -> dict[str, Any]:
+    """Classify a GET-branch HTTP status into a secret-free readback result.
+
+    R1: A bare/generic/repository/wrong-host 404 never yields
+    ``verified_absent=True``. Only an authoritative *branch-scoped* not-found
+    (``not_found_scope='branch'``) may verify deletion.
+    """
+    if status_code == 404:
+        scope = _norm_str(not_found_scope).lower() or NOT_FOUND_SCOPE_UNKNOWN
+        if scope == NOT_FOUND_SCOPE_BRANCH:
+            return {
+                "status": READBACK_NOT_FOUND,
+                "error_class": None,
+                "verified_absent": True,
+                "branch_present": False,
+                "not_found_scope": NOT_FOUND_SCOPE_BRANCH,
+                "reasons": [],
+            }
+        # repository / host / unknown / generic 404 — not verified absence
+        reason = {
+            NOT_FOUND_SCOPE_REPOSITORY: (
+                "post-delete readback 404 is repository-scoped, not branch absence"
+            ),
+            NOT_FOUND_SCOPE_HOST: (
+                "post-delete readback 404 is host-scoped, not branch absence"
+            ),
+        }.get(
+            scope,
+            "post-delete readback 404 is ambiguous (not branch-scoped); "
+            "cannot verify absence",
+        )
+        return {
+            "status": READBACK_AMBIGUOUS_404,
+            "error_class": ERROR_CLASS_UNEXPECTED,
+            "verified_absent": False,
+            "branch_present": None,
+            "not_found_scope": scope,
+            "reasons": [reason],
+        }
+    if status_code in (401, 407):
+        return {
+            "status": READBACK_AUTHENTICATION,
+            "error_class": ERROR_CLASS_AUTHENTICATION,
+            "verified_absent": False,
+            "branch_present": None,
+            "reasons": ["post-delete branch readback authentication failed"],
+        }
+    if status_code == 403:
+        return {
+            "status": READBACK_AUTHORIZATION,
+            "error_class": ERROR_CLASS_AUTHORIZATION,
+            "verified_absent": False,
+            "branch_present": None,
+            "reasons": ["post-delete branch readback authorization failed"],
+        }
+    if status_code is not None and 200 <= int(status_code) < 300:
+        return {
+            "status": READBACK_EXISTS,
+            "error_class": None,
+            "verified_absent": False,
+            "branch_present": True,
+            "reasons": ["post-delete readback found branch still present"],
+        }
+    if status_code is not None and int(status_code) >= 500:
+        return {
+            "status": READBACK_TRANSPORT,
+            "error_class": ERROR_CLASS_TRANSPORT,
+            "verified_absent": False,
+            "branch_present": None,
+            "reasons": ["post-delete branch readback transport/upstream failure"],
+        }
+    return {
+        "status": READBACK_UNEXPECTED,
+        "error_class": ERROR_CLASS_UNEXPECTED,
+        "verified_absent": False,
+        "branch_present": None,
+        "reasons": ["post-delete branch readback returned an unexpected response"],
+        "http_status": status_code,
+    }
+
+
+def _extract_http_status(exc: BaseException) -> int | None:
+    """Extract an HTTP status code from an exception chain (secret-free)."""
+    seen: set[int] = set()
+    current: BaseException | None = exc
+    while current is not None and id(current) not in seen:
+        seen.add(id(current))
+        for attr in ("code", "status", "status_code"):
+            value = getattr(current, attr, None)
+            if isinstance(value, int) and 100 <= value <= 599:
+                return value
+        text = str(current) if current is not None else ""
+        match = re.match(r"HTTP\s+(\d{3})\b", text)
+        if match:
+            return int(match.group(1))
+        current = current.__cause__ or current.__context__
+    return None
+
+
+def classify_branch_readback_exception(
+    exc: BaseException,
+    *,
+    not_found_scope: str | None = None,
+) -> dict[str, Any]:
+    """Classify a GET-branch exception without leaking credentials or bodies.
+
+    R1: substring ``404`` / ``not found`` alone never becomes verified_absent.
+    Callers must pass ``not_found_scope='branch'`` only after authoritative
+    proof that the repository/host is still reachable and the 404 is branch-level.
+    """
+    status_code = _extract_http_status(exc)
+    if status_code is None:
+        lower = str(exc).lower() if exc is not None else ""
+        if any(
+            token in lower
+            for token in (
+                "timed out",
+                "timeout",
+                "connection",
+                "network",
+                "temporarily unavailable",
+                "name or service not known",
+                "nodename nor servname",
+            )
+        ):
+            return {
+                "status": READBACK_TRANSPORT,
+                "error_class": ERROR_CLASS_TRANSPORT,
+                "verified_absent": False,
+                "branch_present": None,
+                "reasons": [
+                    "post-delete branch readback transport/upstream failure"
+                ],
+            }
+        if "unauthorized" in lower:
+            status_code = 401
+        elif "forbidden" in lower:
+            status_code = 403
+        elif "404" in lower or "not found" in lower:
+            # Ambiguous: do NOT treat as branch absence without scope proof.
+            status_code = 404
+            if not_found_scope is None:
+                not_found_scope = NOT_FOUND_SCOPE_UNKNOWN
+
+    if status_code is not None:
+        return classify_branch_readback_http_status(
+            status_code, not_found_scope=not_found_scope
+        )
+
+    return {
+        "status": READBACK_UNEXPECTED,
+        "error_class": ERROR_CLASS_UNEXPECTED,
+        "verified_absent": False,
+        "branch_present": None,
+        "reasons": ["post-delete branch readback returned an unexpected response"],
+    }
+
+
+def assess_post_delete_readback(readback: dict[str, Any] | None) -> dict[str, Any]:
+    """Decide whether DELETE success may be reported after branch readback.
+
+    Success requires authoritative branch-scoped not-found
+    (``verified_absent=True``). DELETE HTTP success alone is never enough.
+    Generic/repository/host 404 cannot verify absence (R1).
+    """
+    payload = dict(readback or {})
+    status = _norm_str(payload.get("status")) or READBACK_UNEXPECTED
+    # Only explicit verified_absent flag counts — never infer from status alone
+    # when status is a bare not_found without branch scope proof.
+    verified = bool(payload.get("verified_absent")) is True
+    if verified and status == READBACK_NOT_FOUND:
+        return {
+            "ok": True,
+            "success": True,
+            "verified_absent": True,
+            "readback": {
+                "status": READBACK_NOT_FOUND,
+                "verified_absent": True,
+                "branch_present": False,
+                "error_class": None,
+                "not_found_scope": NOT_FOUND_SCOPE_BRANCH,
+            },
+            "reasons": [],
+        }
+
+    reasons = list(payload.get("reasons") or [])
+    if not reasons:
+        if status == READBACK_EXISTS:
+            reasons = ["post-delete readback found branch still present"]
+        elif status == READBACK_AUTHENTICATION:
+            reasons = ["post-delete branch readback authentication failed"]
+        elif status == READBACK_AUTHORIZATION:
+            reasons = ["post-delete branch readback authorization failed"]
+        elif status == READBACK_TRANSPORT:
+            reasons = ["post-delete branch readback transport/upstream failure"]
+        elif status == READBACK_AMBIGUOUS_404:
+            reasons = [
+                "post-delete readback 404 is not branch-scoped; "
+                "cannot verify absence"
+            ]
+        else:
+            reasons = ["post-delete branch readback could not verify deletion"]
+
+    return {
+        "ok": False,
+        "success": False,
+        "verified_absent": False,
+        "readback": {
+            "status": status,
+            "verified_absent": False,
+            "branch_present": payload.get("branch_present"),
+            "error_class": payload.get("error_class"),
+            "not_found_scope": payload.get("not_found_scope"),
+        },
+        "reasons": reasons,
+        "blocker_kind": "post_delete_readback_failed",
+    }
+
+
+def cleanup_result_envelope(
+    *,
+    success: bool,
+    performed: bool,
+    delete_acknowledged: bool,
+    verified_absent: bool,
+    **extra: Any,
+) -> dict[str, Any]:
+    """R2: consistent top-level cleanup result fields on every return path."""
+    out: dict[str, Any] = {
+        "success": bool(success),
+        "performed": bool(performed),
+        "delete_acknowledged": bool(delete_acknowledged),
+        "verified_absent": bool(verified_absent),
+    }
+    out.update(extra)
+    return out
+
+
+def _repo_matches(
+    record: dict[str, Any],
+    *,
+    remote: str,
+    org: str,
+    repo: str,
+    host: str | None = None,
+) -> bool:
+    """Match ownership record to target remote/org/repo and normalized host."""
+    if (
+        _norm_str(record.get("remote")).lower() != _norm_str(remote).lower()
+        or _norm_str(record.get("org")).lower() != _norm_str(org).lower()
+        or _norm_str(record.get("repo")).lower() != _norm_str(repo).lower()
+    ):
+        return False
+    expected_host = normalize_host(host)
+    record_host = normalize_host(record.get("host") or record.get("host_name"))
+    # When both sides declare a host, they must agree after normalization.
+    # A record host that disagrees with the expected host is out of scope.
+    # Legacy records without host still match when remote/org/repo agree.
+    if expected_host and record_host and expected_host != record_host:
+        return False
+    return True
+
+
+def _branch_matches(record: dict[str, Any], branch: str) -> bool:
+    return _norm_str(record.get("branch")) == _norm_str(branch)
+
+
+def assess_ownership_record_activity(record: dict[str, Any]) -> dict[str, Any]:
+    """Classify one ownership record as blocking or non-blocking.
+
+    Distinguishes active ownership from expired/released/terminal/stale records.
+    Expired/stale records block unless reclaim_allowed is *explicitly* True
+    (O2: never treat missing/unknown reclaim as auto-allowed).
+    """
+    status = _norm_str(record.get("status")).lower()
+    category = _norm_str(record.get("category")) or "unknown"
+    reclaim_allowed = record.get("reclaim_allowed")
+
+    if category == OWNERSHIP_CATEGORY_INVENTORY_ERROR:
+        return {
+            "blocks": True,
+            "status": status or "unknown",
+            "category": category,
+            "reason": "ownership inventory failed closed",
+        }
+
+    if status in _TERMINAL_OWNERSHIP_STATUSES:
+        return {
+            "blocks": False,
+            "status": status,
+            "category": category,
+            "reason": f"{category} ownership is terminal/released ({status})",
+        }
+    if status in _ACTIVE_OWNERSHIP_STATUSES:
+        return {
+            "blocks": True,
+            "status": status,
+            "category": category,
+            "reason": f"active {category} ownership still uses the target branch",
+        }
+    if status in _EXPIRED_STATUSES or status in _STALE_STATUSES:
+        # O2: only explicit reclaim_allowed=True skips the block.
+        if reclaim_allowed is True:
+            return {
+                "blocks": False,
+                "status": status,
+                "category": category,
+                "reason": (
+                    f"{category} ownership is {status} and reclaimable; "
+                    "does not block deletion"
+                ),
+            }
+        return {
+            "blocks": True,
+            "status": status,
+            "category": category,
+            "reason": (
+                f"{status} {category} ownership still protects the target "
+                "branch (reclaim not proven; fail closed)"
+            ),
+        }
+    # Unknown status → fail closed
+    return {
+        "blocks": True,
+        "status": status or "unknown",
+        "category": category,
+        "reason": (
+            f"unclassified {category} ownership status "
+            f"'{status or 'unknown'}'; fail closed"
+        ),
+    }
+
+
+def assess_active_branch_ownership(
+    *,
+    remote: str,
+    org: str,
+    repo: str,
+    branch: str,
+    host: str | None = None,
+    records: list[dict[str, Any]] | None = None,
+) -> dict[str, Any]:
+    """Assess whether any active ownership still uses *branch* in *repo*.
+
+    Matching requires remote/org/repo/branch and, when provided, normalized
+    host identity. Other repositories, hosts, or branches never false-block.
+    """
+    target_branch = _norm_str(branch)
+    expected_host = normalize_host(host)
+    considered: list[dict[str, Any]] = []
+    blocking: list[dict[str, Any]] = []
+    ignored: list[dict[str, Any]] = []
+
+    for raw in records or []:
+        if not isinstance(raw, dict):
+            continue
+        if not _repo_matches(
+            raw, remote=remote, org=org, repo=repo, host=expected_host or None
+        ):
+            ignored.append(
+                {
+                    "category": _norm_str(raw.get("category")) or "unknown",
+                    "reason": "different repository or host scope",
+                }
+            )
+            continue
+        if not _branch_matches(raw, target_branch):
+            ignored.append(
+                {
+                    "category": _norm_str(raw.get("category")) or "unknown",
+                    "reason": "different branch",
+                }
+            )
+            continue
+        activity = assess_ownership_record_activity(raw)
+        entry = {
+            "category": activity["category"],
+            "status": activity["status"],
+            "blocks": activity["blocks"],
+            "reason": activity["reason"],
+        }
+        considered.append(entry)
+        if activity["blocks"]:
+            blocking.append(entry)
+
+    block = bool(blocking)
+    categories = sorted({b["category"] for b in blocking})
+    reasons = [
+        (
+            "active ownership protects branch "
+            f"'{target_branch}': " + "; ".join(b["reason"] for b in blocking)
+        )
+    ] if block else []
+    return {
+        "block": block,
+        "safe_to_delete": not block,
+        "remote": remote,
+        "org": org,
+        "repo": repo,
+        "host": expected_host or None,
+        "branch": target_branch,
+        "blocking_categories": categories,
+        "blocking": blocking,
+        "considered": considered,
+        "ignored_out_of_scope": ignored,
+        "reasons": reasons,
+        "blocker_kind": "active_branch_ownership" if block else None,
+        "recommended_action": "keep_remote_branch" if block else "delete_remote_branch",
+    }

docs/architecture/mcp-stable-control-runtime-policy-adr.md

@@ -0,0 +1,198 @@
+# ADR: Stable control runtime vs dev runtime (Gitea MCP)
+
+- **Status:** Accepted (policy effective immediately for LLM sessions; tooling may lag)
+- **Date:** 2026-07-09
+- **Tracking issue:** [#615](https://gitea.prgs.cc/Scaled-Tech-Consulting/Gitea-Tools/issues/615)
+- **Related:**
+  - [#543](https://gitea.prgs.cc/Scaled-Tech-Consulting/Gitea-Tools/issues/543) / `docs/mcp-namespace-health.md` — client-namespace health
+  - `docs/mcp-namespace-eof-recovery.md` — reconnect-only EOF recovery (no PID kill)
+  - [#558](https://gitea.prgs.cc/Scaled-Tech-Consulting/Gitea-Tools/issues/558) / `docs/mcp-daemon-import-guard.md` — sanctioned daemon
+  - [#557](https://gitea.prgs.cc/Scaled-Tech-Consulting/Gitea-Tools/issues/557) / `docs/bootstrap-review-path.md` — controller bootstrap for self-hosted fixes
+  - Allocator / control-plane ADR: `docs/architecture/mcp-allocator-control-plane-observability-adr.md` (#613 / PR #614)
+
+## 1. Context
+
+The Gitea MCP server is the **control plane** for real issue/PR mutations (create, comment, lock, review, merge, etc.). When author/reviewer/merger/reconciler sessions kill or restart that process, relaunch it from a feature worktree, or edit the checkout that process loads, operators observe:
+
+- Mid-session identity/preflight resets
+- Stale-runtime vs master parity failures
+- IDE transport EOF / “tool not found” while code on disk has changed
+- Accidental production mutations from experimental code
+
+This ADR separates **stable control runtime** from **dev/test runtime** and defines promotion proof.
+
+## 2. Decision
+
+### 2.1 Stable control runtime
+
+The Gitea MCP server used for **real workflow mutations** is the **stable control runtime**.
+
+Characteristics:
+
+- Loads a known, promoted revision of Gitea-Tools (or the packaged release layout operators designate)
+- Registered in the IDE/client as the production namespaces (`gitea-tools`, `gitea-reviewer`, `gitea-merger`, `gitea-reconciler`, etc.)
+- Holds production profile credentials via sanctioned keychain/env paths only
+
+### 2.2 Dev / test runtime
+
+MCP **server code** development and testing:
+
+- Happens in isolated **`branches/`** worktrees (or other non-stable checkouts)
+- May use a **separate** dev/test MCP runtime/process when process-level testing is required
+- **Must not** be used for real Gitea mutations on production issues/PRs
+
+### 2.3 Forbidden actions (normal sessions)
+
+Normal **author, reviewer, merger, and reconciler** LLM sessions **must not**:
+
+| Forbidden | Why |
+|-----------|-----|
+| Kill the running MCP server process | Drops all concurrent sessions; loses preflight state |
+| Restart / relaunch the MCP server process | Same as kill; causes stale/identity churn mid-workflow |
+| Relaunch MCP from a development worktree | Runs unpromoted code against production mutations |
+| Edit files in the stable runtime checkout | Hot-mutates control plane under concurrent users |
+| Use experimental/dev MCP for real Gitea mutations | Bypasses promotion proof and audit expectations |
+| Bypass or self-reset a stale master-parity gate | The gate is fail-closed; only an operator reload restores parity |
+
+**LLM-allowed vs operator-owned (authoritative split):**
+
+| Actor | May do | Must not do |
+|-------|--------|-------------|
+| **LLM session** | Call tools on the already-running stable namespaces; **client reconnect** after transport EOF (no process kill); pass `worktree_path` / role worktree args; report blockers and stop mutations when unhealthy/stale | Kill, restart, or relaunch any MCP process; bump config mtimes to force reload; edit the stable checkout; switch to a dev MCP for production mutations |
+| **Operator / release-manager** | Supervised restart/reload of the **stable** control runtime; dual-namespace client configuration; §2.4 promotions; incident recovery | — |
+
+EOF / transport recovery for LLM sessions: **client reconnect only** (see `docs/mcp-namespace-eof-recovery.md`). Do not “fix” health by killing PIDs or bumping MCP config mtimes as a normal session procedure.
+
+This ADR **supersedes** any older runbook wording that told the LLM to relaunch or restart the client/MCP as a self-service step. Where `docs/llm-workflow-runbooks.md` (or wiki runbooks) discuss dual-namespace setup or workspace rebind, **process restart/relaunch is operator-owned**; the LLM stops, reports, and waits.
+
+### 2.4 Promotion (operator / release-manager only)
+
+Promotion of a new revision into the stable control runtime is an **explicit operator/release-manager action**, not an LLM self-service step.
+
+A promotion **must record** (issue comment, release note, or promotion ledger):
+
+| Field | Description |
+|-------|-------------|
+| **previous runtime SHA** | Commit previously loaded by stable runtime |
+| **promoted runtime SHA** | Commit after promotion |
+| **source branch/PR** | Where the change was reviewed |
+| **restart/reload method** | How the process was cycled (e.g. supervised restart, client reload) |
+| **health check proof** | Client-namespace probe success (`gitea_whoami` / namespace health) |
+| **identity/profile proof** | Expected profile(s) and username(s) after reload |
+| **workspace/root proof** | Stable checkout path / root matches intended layout |
+| **mutation capability proof** | Required permissions for the target role present; forbidden ops still forbidden |
+| **rollback instructions** | How to restore previous SHA and re-verify health |
+
+Suggested durable marker:
+
+```text
+## MCP STABLE RUNTIME PROMOTION (#615)
+
+Status: COMPLETED | ROLLED_BACK | ABORTED
+Previous-SHA: <full sha>
+Promoted-SHA: <full sha>
+Source-PR: <number>
+Source-Branch: <name>
+Reload-Method: <text>
+Health-Proof: client_namespace whoami OK / assess_mcp_namespace_health OK
+Identity-Proof: profile=<name> user=<name>
+Workspace-Proof: root=<path>
+Mutation-Proof: allowed_ops include <…>; forbidden include <…>
+Rollback: checkout <previous sha>; reload method <…>; re-run health/identity proofs
+Operator: <username>
+Timestamp: <ISO-8601>
+```
+
+### 2.5 Unhealthy stable runtime → stop work
+
+If the stable MCP runtime is **unhealthy**, including any of:
+
+- client-namespace probes fail
+- wrong identity / wrong profile
+- wrong workspace root
+- missing mutation capability for the intended role
+- persistent EOF after **client reconnect**
+- **master parity is stale** (`startup_head` behind on-disk `master` / `restart_required` from the parity gate)
+
+then:
+
+1. **Normal PR / review / merge / issue-mutation work must stop immediately.**
+2. The session **reports** the unhealthy/stale state (tool error, CTH, or operator handoff) with startup vs current head when known.
+3. Do **not** improvise: no LLM process kill/restart, no dev-worktree MCP for production mutations, no env escape hatches, no manual gate bypass.
+4. Resume only after:
+   - an **operator** restores the runtime (see §2.6 for routine post-merge parity reload), **or**
+   - a **controlled** promotion/rollback completes with the §2.4 promotion record, **or**
+   - a controller invokes the narrow **bootstrap review path** (#557) when the defect is self-hosted and documented,
+   - **and** the session re-verifies health (and master parity when applicable) before the next mutation.
+
+### 2.6 Routine post-merge master-parity staleness (operator reload, not promotion)
+
+**Symptom:** After merges land on `master`, a long-lived stable MCP process still runs the pre-merge `startup_head`. The master-parity gate marks the server **stale** / `restart_required` and **blocks mutations**.
+
+**Sanctioned response (authoritative):**
+
+| Step | Actor | Action |
+|------|-------|--------|
+| 1 | LLM session | Mutations stop immediately when the gate reports stale. |
+| 2 | LLM session | Report the stale state (startup head, current master head, that operator reload is required). Do not retry mutations. |
+| 3 | **Operator** | Reload/restart the **stable** control MCP so it loads current `master` (supervised client/daemon reload). |
+| 4 | LLM session | Resume only after startup/current-head parity is verified (e.g. `gitea_get_runtime_context` / parity assessment shows in parity). |
+
+**Not a §2.4 promotion:** Catching the already-designated stable control checkout up to a newly advanced `master` is a **routine operator reload** of the stable runtime. It does **not** require the nine-field promotion ledger. Use §2.4 only when **changing which unpromoted/dev revision** becomes the stable control runtime (new source branch/PR into the stable designation).
+
+**Code note:** `master_parity_gate.py` may still say “restart the server” in machine-facing reason strings. That string names the **operator recovery action**, not an LLM self-service instruction. This ADR and the runbooks define the actor split.
+
+## 3. Relationship to other controls
+
+| Doc / mechanism | Interaction |
+|-----------------|-------------|
+| Namespace health (#543) | Proves IDE client can call tools; does not authorize restart |
+| EOF recovery | Reconnect only; no process kill |
+| Daemon import guard (#558) | Mutations require sanctioned daemon; not a bare shell import |
+| Bootstrap path (#557) | Only controller-authorized exception when live runtime cannot review its own fix |
+| Allocator / control-plane ADR | Coordination DB is separate; still depends on a healthy MCP surface for Gitea writes |
+
+## 4. Consequences
+
+### Positive
+
+- Predictable control plane for concurrent LLMs
+- Clear operator-only promotion gate with rollback
+- Aligns session behavior with health/EOF docs already landed
+
+### Costs
+
+- LLM sessions must wait when runtime is sick (no DIY restart)
+- Operators must maintain promotion discipline and dual-runtime config if they use a dev MCP
+
+### Non-goals
+
+- Does not ban operator-supervised restarts during incidents
+- Does not replace CI or code review for MCP changes
+- Does not authorize editing stable checkout “because tests need a quick fix”
+
+## 5. Implementation follow-ups (optional tooling)
+
+These may land in later issues; the **policy binds sessions now**:
+
+1. Session preflight that refuses mutations if workspace root equals a `branches/` feature worktree configured as “dev only.”
+2. Explicit `runtime_kind=stable|dev` in MCP config and `gitea_whoami` profile metadata.
+3. Promotion checklist script that emits the durable promotion marker fields.
+
+**Not optional (issue #615 acceptance criterion 2):** operator guide and runbooks **must** cross-link this ADR (see §6). Cross-links are documentation acceptance, not deferred tooling.
+
+## 6. Acceptance for this ADR
+
+1. Document merged under `docs/architecture/mcp-stable-control-runtime-policy-adr.md`.
+2. **Operator guide / runbooks cross-link this ADR** (`docs/wiki/Operator-Guide.md`, `docs/wiki/Runbooks.md`, `docs/llm-workflow-runbooks.md`).
+3. Issue #615 references this path.
+4. LLM/operator runbooks treat kill/restart/relaunch-from-worktree as **LLM violations**; process restart is **operator-owned**.
+5. Unhealthy runtime (including **stale master parity**) stops normal mutation work until operator restore/reload, promotion/rollback, or #557 bootstrap — then re-verify parity before mutating.
+6. Routine post-merge parity reload is documented as operator reload (§2.6), not an LLM self-restart and not a full §2.4 promotion.
+
+## 7. Document history
+
+| Date | Change |
+|------|--------|
+| 2026-07-09 | Initial ADR: stable vs dev runtime, forbidden session actions, promotion proof fields, stop-work rule |
+| 2026-07-16 | Review 443 remediation (#615 / PR #616): mandatory cross-links; LLM vs operator restart split; routine post-merge parity staleness (§2.6); stale parity in §2.5 unhealthy triggers |

docs/gitea-execution-profiles.md

@@ -238,11 +238,43 @@ narrow operation set:
 - `gitea.issue.comment`
 - `gitea.issue.close`
 - `gitea.pr.close`
+- `gitea.branch.delete` (merged-branch cleanup only — see below)
 
 Forbidden on reconciler profiles: `gitea.pr.approve`, `gitea.pr.merge`,
 `gitea.pr.review`, `gitea.pr.create`, `gitea.branch.push`, and
 `gitea.repo.commit`.
 
+### Merged-branch cleanup ownership (`gitea.branch.delete`)
+
+The reconciler is the repository-supported owner of merged-PR source-branch
+cleanup: `task_capability_map` maps `cleanup_merged_pr_branch` (and
+`reconciliation_cleanup`) to role `reconciler` with permission
+`gitea.branch.delete`. Post-merge branch lifecycle is reconciliation work —
+it happens after the author, reviewer, and merger roles have completed, and
+it must not be reachable from those roles.
+
+Least-privilege constraints:
+
+- `gitea.branch.delete` is granted **only** to reconciler profiles. Author,
+  reviewer, and merger profiles must never hold it; `gitea_delete_branch`
+  and `gitea_cleanup_merged_pr_branch` fail closed on any profile without
+  the permission.
+- Even with the permission, reconciler deletion is only supported through the
+  guarded `gitea_cleanup_merged_pr_branch` path (#514 / #687): the PR must be
+  merged, the head an ancestor of the target, the branch not protected
+  (`master`/`main`/`dev`), the branch not a preservation/evidence ref (e.g.
+  `chore/issue-681-preserve-review-session-wip`), no open PR may still use the
+  head, and an explicit `CLEANUP MERGED PR <n> BRANCH <branch>` confirmation is
+  required. Raw `gitea_delete_branch` is **denied** to reconciler even when
+  `gitea.branch.delete` is present.
+- Raw `git branch -d` / `git push --delete` cleanup remains blocked by
+  `branch_cleanup_guard` and the final-report validator regardless of
+  profile permissions.
+- `gitea.branch.delete` has no short alias in `GITEA_OPERATION_ALIASES`;
+  write it fully qualified in `allowed_operations`. Migration must emit
+  canonical names such as `gitea.pr.close` (never bare `pr.close` /
+  `issue.close`, which the production normalizer rejects or drops).
+
 Launch a static `gitea-reconciler` MCP namespace with
 `GITEA_MCP_PROFILE=prgs-reconciler`. Profile shape is validated by
 `reconciler_profile.assess_reconciler_profile` (#304). Use the
@@ -251,6 +283,159 @@ Launch a static `gitea-reconciler` MCP namespace with
 fresh target-branch fetch, recorded target SHA, and ancestor proof. PRs whose
 heads are not already landed cannot be closed through this path.
 
+### Operational runbook: grant reconciler `gitea.branch.delete` (#687)
+
+Merging a code PR that updates `migrate_profiles.py` / `reconciler_profile.py`
+**does not** change the live operator profile on disk. Apply the profile
+change deliberately, then reconnect the client-managed namespace.
+
+1. **Approved migration / profile-update command** (from the repo root, using
+   the project venv if present):
+
+   ```bash
+   # Dry-run first (default): validates v2 output, writes nothing
+   python3 migrate_profiles.py -i ~/.config/gitea-tools/profiles.json
+
+   # Apply: creates backup then writes migrated v2 config
+   python3 migrate_profiles.py -i ~/.config/gitea-tools/profiles.json -w
+   # Optional explicit paths:
+   # python3 migrate_profiles.py -i ~/.config/gitea-tools/profiles.json \
+   #   -o ~/.config/gitea-tools/profiles.json \
+   #   --backup ~/.config/gitea-tools/profiles.json.bak -w
+   ```
+
+   If the live file is already v2, edit the reconciler identity’s
+   `allowed_operations` / `forbidden_operations` under
+   `environments.<env>.services.gitea.identities.reconciler` (or the
+   `prgs-reconciler` alias target) so allowed includes the canonical set
+   below — then re-validate with a load of the config (see step 3).
+
+2. **Inspect the generated (or edited) profile** — confirm the reconciler
+   identity, for example:
+
+   ```bash
+   python3 - <<'PY'
+   import json
+   from pathlib import Path
+   cfg = json.loads(Path.home().joinpath(".config/gitea-tools/profiles.json").read_text())
+   # v2 environments shape:
+   ident = cfg["environments"]["prgs"]["services"]["gitea"]["identities"]["reconciler"]
+   print("role:", ident.get("role"))
+   print("allowed:", ident.get("allowed_operations"))
+   print("forbidden:", ident.get("forbidden_operations"))
+   PY
+   ```
+
+3. **Validate canonical operation names and least privilege**
+
+   Expected canonical **allowed** (defaults after migration):
+
+   - `gitea.read`
+   - `gitea.pr.close` (required)
+   - `gitea.pr.comment`
+   - `gitea.issue.comment`
+   - `gitea.issue.close`
+   - `gitea.branch.delete` (recommended; cleanup only)
+
+   Expected **forbidden** includes at least: `gitea.pr.approve`,
+   `gitea.pr.merge`, `gitea.pr.review`, `gitea.pr.create`,
+   `gitea.branch.push`, `gitea.repo.commit`.
+
+   No shorthand (`pr.close`, `issue.close`, `pr.comment`) may remain.
+   Validate with the production loader:
+
+   ```bash
+   python3 - <<'PY'
+   import gitea_config, reconciler_profile
+   from pathlib import Path
+   path = str(Path.home() / ".config/gitea-tools/profiles.json")
+   gitea_config.load_config(path)  # fails closed on invalid config
+   # Or assess the reconciler lists directly after extracting them:
+   # print(reconciler_profile.assess_reconciler_profile(allowed, forbidden))
+   PY
+   ```
+
+4. **Merging PR #688 (or any code PR) does not update the live profile.**
+   Code changes only the migration helper, schema, docs, and tests. The
+   operator must still run `migrate_profiles.py -w` or an equivalent
+   authorized edit of `~/.config/gitea-tools/profiles.json`.
+
+5. **Supported apply method:** `python3 migrate_profiles.py … -w` (backup
+   created automatically) **or** operator-authorized edit of the live
+   profiles file after backup. Unsupported: silent mtime tricks, manual
+   process kill to “reload”, or undocumented env overrides.
+
+6. **Backup and validation:** `-w` copies the input to
+   `<input_path>.bak` (or `--backup PATH`) before writing. Re-run
+   `load_config` / `assess_reconciler_profile` after write. Keep the
+   `.bak` until live whoami/capability checks pass.
+
+7. **Client-managed namespace reconnect/reload:** reconnect or reload the
+   IDE MCP client so `gitea-reconciler` restarts from current `master` and
+   the updated `GITEA_MCP_PROFILE=prgs-reconciler` config. Do not hand-launch
+   `mcp_server.py` / `gitea_mcp_server.py` with ad hoc `GITEA_*` env
+   (see #686 / #630).
+
+8. **Live reverification** (through the client-managed `gitea-reconciler`
+   namespace only):
+
+   - `gitea_whoami` → identity + profile `prgs-reconciler`
+   - `gitea_assess_master_parity` → `stale=false`, `restart_required=false`
+   - `gitea_resolve_task_capability(task="cleanup_merged_pr_branch")` →
+     `allowed_in_current_session=true` only when permission and role match
+   - `gitea_resolve_task_capability(task="delete_branch")` →
+     **not** allowed for reconciler (role denial must be enforced)
+
+9. **Guarded cleanup usage** (example for a merged PR whose source branch
+   remains on the remote):
+
+   ```text
+   gitea_cleanup_merged_pr_branch(
+     pr_number=<N>,
+     branch=<exact PR head branch>,
+     confirmation="CLEANUP MERGED PR <N> BRANCH <exact PR head branch>",
+     remote="prgs",
+     org="Scaled-Tech-Consulting",
+     repo="Gitea-Tools",
+     worktree_path="<path under branches/>",
+   )
+   ```
+
+   The tool refuses unmerged PRs, protected branches, preservation/evidence
+   branches, open-PR heads, mismatched branch names, and wrong confirmation.
+
+10. **Prohibitions**
+
+    - No raw `git push --delete`, `git branch -d` / `-D`, or delete refspecs
+    - No arbitrary `gitea_delete_branch` from reconciler
+    - No unsupported profile switching mid-run without full re-preflight
+    - No ad hoc hand-edits of live profiles **unless** operator-authorized,
+      backed up, and revalidated as above
+
+Canonical migrated reconciler example:
+
+```json
+{
+  "role": "reconciler",
+  "allowed_operations": [
+    "gitea.read",
+    "gitea.pr.close",
+    "gitea.pr.comment",
+    "gitea.issue.comment",
+    "gitea.issue.close",
+    "gitea.branch.delete"
+  ],
+  "forbidden_operations": [
+    "gitea.pr.approve",
+    "gitea.pr.merge",
+    "gitea.pr.review",
+    "gitea.pr.create",
+    "gitea.branch.push",
+    "gitea.repo.commit"
+  ]
+}
+```
+
 ## Identity and fail-closed rules
 
 Before **any** mutating action, a workflow must know both:

docs/llm-workflow-runbooks.md

@@ -195,7 +195,7 @@ To avoid the bottleneck of relaunching/restarting the MCP server to switch betwe
   `gitea_reconcile_already_landed_pr` after ancestry proof — not for normal
   review or author workflows.
 
-* **Fallback:** If the dual-profile MCP launcher pattern is not supported or configured in the client, the LLM must relaunch or restart the client/MCP with the correct profile environment variable before claiming or working on any tasks.
+* **Fallback (operator-owned):** If the dual-profile MCP launcher pattern is not supported or configured in the client, **do not** have the LLM relaunch or restart the client/MCP. The LLM **stops** role-switching work, reports that the correct static namespace is missing, and waits for an **operator** to configure dual namespaces or reload the client with the correct `GITEA_MCP_PROFILE` for that role. Process restart/relaunch is operator-owned under the [stable control runtime ADR](architecture/mcp-stable-control-runtime-policy-adr.md) (#615).
 
 ## Setup runbook — interactive menu
 
@@ -1200,10 +1200,13 @@ When a mutation blocks on workspace binding:
 
 1. Read the error — it names the **resolved workspace path**, **role
    namespace**, and **binding source** (tool arg, env var, or process root).
-2. Reconnect or relaunch the correct namespace MCP server from the intended
+2. **LLM-allowed:** pass `worktree_path` on reviewer/merger mutation tools when
-   workspace (or set the role-specific env var before launch).
+   the active `branches/` worktree differs from the MCP process root; **client
-3. Pass `worktree_path` on reviewer/merger mutation tools when the active
+   reconnect** after transport EOF only (no process kill).
-   branches/ worktree differs from the MCP process root.
+3. **Operator-owned:** if the wrong namespace process was launched, or a role-
+   specific `GITEA_*_WORKTREE` must be set at process start, an **operator**
+   reloads/relaunches the correct static namespace MCP. LLM sessions must not
+   kill or restart MCP processes (see [stable control runtime ADR](architecture/mcp-stable-control-runtime-policy-adr.md)).
 4. **Do not** clean, reset, or discard foreign role worktrees to unblock your
    own namespace — that destroys another agent's WIP.
 
@@ -1213,9 +1216,10 @@ When posting a Canonical Thread Handoff after a binding blocker:
 
 - State which namespace was active (author / reviewer / merger / reconciler).
 - Quote the resolved workspace path and binding source from the error.
-- Name the safe reconnect action (relaunch MCP from `branches/...`, set
+- Name the safe next action: pass `worktree_path` if that unblocks the tool, or
-  `GITEA_*_WORKTREE`, or pass `worktree_path`).
+  request an **operator** reload of the correct namespace MCP / env binding.
-- Explicitly note that foreign worktrees must not be cleaned to unblock.
+- Explicitly note that foreign worktrees must not be cleaned to unblock, and
+  that the LLM must not self-restart the MCP process.
 
 ## Safety notes
 
@@ -1226,6 +1230,7 @@ When posting a Canonical Thread Handoff after a binding blocker:
 
 ## Related documents
 
+- [`architecture/mcp-stable-control-runtime-policy-adr.md`](architecture/mcp-stable-control-runtime-policy-adr.md) — stable control runtime vs dev runtime; LLM must not kill/restart MCP; operator-owned reload and promotions; routine post-merge parity staleness (#615).
 - [`reviewer-handoff-consistency.md`](reviewer-handoff-consistency.md) — reject contradictory reviewer handoffs (#501).
 - [`issue-acceptance-gate.md`](issue-acceptance-gate.md) — controller issue-acceptance audit after PR merge (#500).
 - [`../skills/llm-project-workflow/SKILL.md`](../skills/llm-project-workflow/SKILL.md) — portable cross-project LLM workflow skill.

docs/wiki/Operator-Guide.md

@@ -14,6 +14,7 @@ Handbook for LLM operators and human developers using the Gitea-Tools MCP server
 4. **No self-review / no self-merge** — The authenticated Gitea user must not approve or merge a PR they authored.
 5. **Follow the gates** — Prompts express intent; MCP tools enforce safety. Never bypass gates via prompt instructions.
 6. **Global LLM Worktree Rule** — Main checkout stays on `master`/`main`/`dev`; all mutations happen under `branches/`. Prove project root, `cwd`, branch, stable main-checkout branch, and session worktree path before editing. No exceptions.
+7. **Stable control runtime** — Real Gitea mutations use only the **stable** MCP control runtime. LLM sessions must not kill, restart, or relaunch MCP processes, edit the stable checkout, or use a dev MCP for production mutations. See the policy ADR: [mcp-stable-control-runtime-policy-adr.md](../architecture/mcp-stable-control-runtime-policy-adr.md) (#615).
 
 ## Supported Gitea instances
 
@@ -30,3 +31,4 @@ Always pass `remote` explicitly on tool calls. The server default is `dadeschool
 2. `gitea_get_runtime_context` — allowed/forbidden operations for this session.
 3. `gitea_resolve_task_capability` — prove the session may perform the planned task.
 4. For reviewer work: dry-run validation (`gitea_dry_run_pr_review`) before live review mutations.
+5. Confirm master parity / runtime health when tools report stale or unhealthy control runtime — stop mutations and request an **operator** reload of the stable MCP (see [stable control runtime ADR](../architecture/mcp-stable-control-runtime-policy-adr.md)).

docs/wiki/Runbooks.md

@@ -12,6 +12,17 @@
 4. `gitea_mark_final_review_decision` → approve via `gitea_review_pr`.
 5. `gitea_merge_pr` with pinned head SHA and `confirmation="MERGE PR <n>"`.
 
+## Stable MCP control runtime (#615)
+
+Policy ADR: [mcp-stable-control-runtime-policy-adr.md](../architecture/mcp-stable-control-runtime-policy-adr.md).
+
+| Situation | Who acts | What to do |
+|-----------|----------|------------|
+| Transport EOF / missing tools | LLM | **Client reconnect only** — do not kill PIDs |
+| Wrong profile / dual-namespace missing | Operator | Configure or reload the correct static namespace(s) |
+| Master parity stale after merge | LLM stops + reports; **operator** reloads stable MCP | Resume only after parity re-verified |
+| Promote unpromoted MCP code to stable | Operator / release-manager only | Full §2.4 promotion record |
+
 ## Gitea Wiki sync
 
 The Gitea Wiki mirrors `docs/wiki/` (source of truth). After merging wiki changes:

gitea_auth.py

@@ -243,6 +243,192 @@ def _redact(text):
         return str(text)
 
 
+# ── Classified client failures (#699) ─────────────────────────────────────────
+# Subclasses of RuntimeError preserve existing ``except RuntimeError`` call
+# sites. Exception *messages* are fixed constants only — HTTP response bodies,
+# Keychain material, and arbitrary exception text are never stored on the
+# exception or re-emitted to tool results / daemon logs.
+
+
+# Fixed messages (must match mcp_tool_error_boundary.FIXED_MESSAGES keys used here).
+_MSG_AUTH_INVALID = "Gitea authentication failed: invalid or revoked credentials"
+_MSG_AUTH_FAILED = "Gitea authentication failed"
+_MSG_AUTHZ_SCOPE = "Gitea authorization failed: insufficient token scope"
+_MSG_AUTHZ_DENIED = "Gitea authorization failed: access denied"
+_MSG_NETWORK = "Network error contacting Gitea"
+_MSG_CONFIG = "Gitea configuration or credential resolution failed"
+_MSG_UPSTREAM = "Gitea upstream unavailable"
+_MSG_HTTP = "Gitea HTTP request failed"
+
+
+class GiteaClientError(RuntimeError):
+    """Base for known Gitea client failures with stable reason_code metadata."""
+
+    reason_code = "client_error"
+    error_class = "client"
+    http_status = None
+
+    def __init__(self, message=None, *, reason_code=None, http_status=None):
+        if reason_code is not None:
+            self.reason_code = reason_code
+        if http_status is not None:
+            self.http_status = http_status
+        # Message is always a fixed constant; callers cannot inject bodies.
+        fixed = message if message is not None else _MSG_HTTP
+        super().__init__(fixed)
+
+
+class GiteaAuthError(GiteaClientError):
+    """Authentication failure (invalid/revoked credentials → typically HTTP 401)."""
+
+    reason_code = "auth_invalid_token"
+    error_class = "authentication"
+    http_status = 401
+
+    def __init__(self, message=None, *, reason_code=None, http_status=None):
+        super().__init__(
+            message if message is not None else _MSG_AUTH_INVALID,
+            reason_code=reason_code or "auth_invalid_token",
+            http_status=http_status if http_status is not None else 401,
+        )
+
+
+class GiteaAuthzError(GiteaClientError):
+    """Authorization failure (HTTP 403 — scope deficiency or access denied)."""
+
+    reason_code = "authz_denied"
+    error_class = "authorization"
+    http_status = 403
+
+    def __init__(self, message=None, *, reason_code=None, http_status=None):
+        code = reason_code or "authz_denied"
+        if code == "authz_insufficient_scope":
+            fixed = _MSG_AUTHZ_SCOPE
+        else:
+            fixed = _MSG_AUTHZ_DENIED
+            code = "authz_denied"
+        super().__init__(
+            message if message is not None else fixed,
+            reason_code=code,
+            http_status=http_status if http_status is not None else 403,
+        )
+
+
+class GiteaNetworkError(GiteaClientError):
+    """Transport / DNS / timeout failure contacting Gitea."""
+
+    reason_code = "network_error"
+    error_class = "network"
+    http_status = None
+
+    def __init__(self, message=None, *, reason_code=None, http_status=None):
+        super().__init__(
+            message if message is not None else _MSG_NETWORK,
+            reason_code=reason_code or "network_error",
+            http_status=http_status,
+        )
+
+
+class GiteaConfigError(GiteaClientError):
+    """Local configuration / credential resolution failure (not HTTP auth)."""
+
+    reason_code = "config_error"
+    error_class = "configuration"
+    http_status = None
+
+    def __init__(self, message=None, *, reason_code=None, http_status=None):
+        super().__init__(
+            message if message is not None else _MSG_CONFIG,
+            reason_code=reason_code or "config_error",
+            http_status=http_status,
+        )
+
+
+class GiteaHttpError(GiteaClientError):
+    """Non-auth HTTP failure with fixed message (no response body)."""
+
+    reason_code = "http_error"
+    error_class = "client"
+    http_status = None
+
+    def __init__(self, message=None, *, reason_code=None, http_status=None):
+        code = reason_code or "http_error"
+        if code == "upstream_unavailable":
+            fixed = _MSG_UPSTREAM
+        else:
+            fixed = _MSG_HTTP
+            code = "http_error"
+        super().__init__(
+            message if message is not None else fixed,
+            reason_code=code,
+            http_status=http_status,
+        )
+
+
+def _looks_like_insufficient_scope(detail: str) -> bool:
+    """Internal: inspect redacted body *only* to refine 403 reason_code.
+
+    The body is never stored on the exception or returned to callers.
+    """
+    lower = (detail or "").lower()
+    markers = (
+        "insufficient scope",
+        "required scope",
+        "does not have at least one of required scope",
+        "token does not have",
+        "missing scope",
+        "scope(s)",
+    )
+    return any(m in lower for m in markers)
+
+
+def classify_http_status(code: int, *, body_hint: str = "") -> tuple[type, str, int]:
+    """Central HTTP status → (exception_class, reason_code, http_status).
+
+    Every HTTP 403 becomes authorization-class. Body text is used only as a
+    local hint for scope vs denied reason_code and is never returned.
+    """
+    if code == 401:
+        return (GiteaAuthError, "auth_invalid_token", 401)
+    if code == 403:
+        if _looks_like_insufficient_scope(body_hint or ""):
+            return (GiteaAuthzError, "authz_insufficient_scope", 403)
+        return (GiteaAuthzError, "authz_denied", 403)
+    if code in (502, 503, 504):
+        return (GiteaHttpError, "upstream_unavailable", code)
+    return (GiteaHttpError, "http_error", code)
+
+
+def raise_for_http_status(code: int, body: str = "") -> None:
+    """Raise a typed client error for *code* without embedding *body*.
+
+    *body* may be inspected only to choose scope vs denied for 403; it is
+    never placed on the exception message.
+    """
+    # Redact before any inspection; discard after classification.
+    try:
+        hint = _redact(body or "").strip()
+    except Exception:
+        hint = ""
+    exc_cls, reason, status = classify_http_status(code, body_hint=hint)
+    # Explicitly construct without passing body/hint into message.
+    if exc_cls is GiteaAuthError:
+        raise GiteaAuthError(reason_code=reason, http_status=status)
+    if exc_cls is GiteaAuthzError:
+        raise GiteaAuthzError(reason_code=reason, http_status=status)
+    if reason == "upstream_unavailable":
+        raise GiteaHttpError(
+            reason_code="upstream_unavailable",
+            http_status=status,
+        )
+    raise GiteaHttpError(reason_code="http_error", http_status=status)
+
+
+def _raise_http_error(code: int, detail: str = "") -> None:
+    """Backward-compatible alias — *detail* is never embedded in the error."""
+    raise_for_http_status(code, detail)
+
+
 def _add_query(url, **params):
     """Return *url* with the given query parameters added or overridden.
 
@@ -315,23 +501,24 @@ def api_request(method, url, auth_header, payload=None, *,
     """Make an authenticated JSON request to the Gitea API.
 
     Returns parsed JSON on success (or ``None`` for an empty body), and raises
-    ``RuntimeError`` on failure.
+    a classified client error on failure.
 
     On HTTP 429 the request is retried up to *max_retries* times: honoring a
     valid ``Retry-After`` header (seconds or HTTP-date) when present, otherwise
     using capped jittered exponential backoff. Successful responses are
     unchanged.
 
-    All failures are converted to a ``RuntimeError`` with a clear, secret
+    Failures raise typed exceptions with **fixed messages only** (#699). HTTP
-    -redacted message (no raw stack traces or credential material):
+    response bodies are read solely for local 403 reason refinement and are
+    never stored on exceptions or returned to callers:
 
-    - Non-429 HTTP errors surface the status code and a redacted response body.
+    - HTTP 401 → :class:`GiteaAuthError` (``auth_invalid_token``)
-      502/503/504 upstream errors get an explicit "Gitea upstream unavailable"
+    - HTTP 403 → :class:`GiteaAuthzError` (scope or denied)
-      message.
+    - 502/503/504 → :class:`GiteaHttpError` (``upstream_unavailable``)
-    - Timeouts and network/DNS failures (``URLError`` / ``TimeoutError``) surface
+    - Other non-429 HTTP → :class:`GiteaHttpError` (``http_error``)
-      a generic "network error contacting Gitea" message.
+    - Timeouts / DNS / ``URLError`` → :class:`GiteaNetworkError`
-    - A malformed (non-JSON) success body surfaces a "malformed JSON response"
+    - Malformed success JSON → plain ``RuntimeError`` (programming/protocol;
-      message rather than a raw decode error.
+      not reclassified as authentication)
 
     The ``*_func`` parameters and ``timeout`` are injection points for
     deterministic testing.
@@ -369,22 +556,23 @@ def api_request(method, url, auth_header, payload=None, *,
                 error_body = e.read().decode("utf-8", errors="replace")
             except Exception:
                 error_body = ""
-            detail = _redact(error_body).strip()
+            # Classify from status (+ local body hint). Body is not embedded.
-            if e.code in (502, 503, 504):
+            try:
-                msg = f"HTTP {e.code}: Gitea upstream unavailable"
+                raise_for_http_status(e.code, error_body)
-                raise RuntimeError(f"{msg}: {detail}" if detail else msg) from e
+            except GiteaClientError:
-            raise RuntimeError(f"HTTP {e.code}: {detail}") from e
+                raise
+            # Defensive: raise_for_http_status always raises.
+            raise GiteaHttpError(http_status=e.code) from e  # pragma: no cover
         except (urllib.error.URLError, TimeoutError) as e:
-            reason = getattr(e, "reason", e)
+            # Fixed message only — do not embed URLError reason (may leak paths).
-            raise RuntimeError(
+            raise GiteaNetworkError(reason_code="network_error") from e
-                f"network error contacting Gitea: {_redact(reason)}"
-            ) from e
 
         if not body:
             return None
         try:
             return json.loads(body)
         except ValueError as e:
+            # Programming/protocol failure — not authentication.
             raise RuntimeError("malformed JSON response from Gitea") from e
 
 

mcp_session_state.py

@@ -36,7 +36,43 @@ KIND_REVIEW_DRAFT = "review_draft"
 # other session proofs; a contaminated session fails closed on gated mutations
 # until a reconciler audits and clears it.
 KIND_STABLE_BRANCH_CONTAMINATION = "stable_branch_contamination"
+# Durable shadow of the in-memory reviewer session lease (#702). Written on
+# every sanctioned record/heartbeat and removed on sanctioned clear, so a
+# daemon that dies without teardown leaves provable orphan evidence (owner
+# pid + session id) for the guarded lease-cleanup path. Never a substitute
+# for the in-session lease: mutation gates ignore it.
+KIND_REVIEWER_SESSION_LEASE = "reviewer_session_lease"
+# Durable audit record written whenever the daemon's sanctioned stale-binding
+# recovery clears an inherited GITEA_ACTIVE_WORKTREE binding (#702).
+KIND_STALE_BINDING_RECOVERY = "stale_binding_recovery"
+# #709: archive prior terminal decision ledgers instead of silent overwrite.
+KIND_DECISION_LOCK_ARCHIVE = "review_decision_lock_archive"
+# #709: post-merge cleanup/audit reconciliation-required durable record.
+KIND_POST_MERGE_DECISION_RECOVERY = "post_merge_decision_recovery"
+# #709: truthful record when historical terminal evidence is irrecoverably gone.
+KIND_IRRECOVERABLE_DECISION_PROVENANCE = "irrecoverable_decision_provenance"
+# #709 F1: server-side non-forgeable authorization artifact for recovery.
+KIND_IRRECOVERABLE_PROVENANCE_AUTH = "irrecoverable_provenance_authorization"
 
+# Kinds that must survive the default session-state TTL (forensic / recovery).
+#
+# KIND_DECISION_LOCK is recovery-critical (#720): terminal review provenance is
+# not disposable cache. A generic four-hour TTL must not drop old-head evidence
+# or make ``fresh_review_on_current_head_allowed`` unreachable. Same-head / same-
+# run #332 protections still apply once the ledger is loadable. Other session
+# kinds (workflow load, drafts, etc.) remain TTL-bound.
+RECOVERY_CRITICAL_KINDS = frozenset(
+    {
+        KIND_DECISION_LOCK,
+        KIND_DECISION_LOCK_ARCHIVE,
+        KIND_POST_MERGE_DECISION_RECOVERY,
+        KIND_IRRECOVERABLE_DECISION_PROVENANCE,
+        KIND_IRRECOVERABLE_PROVENANCE_AUTH,
+        # #702 crash-orphan evidence (must outlive TTL; F4)
+        KIND_REVIEWER_SESSION_LEASE,
+        KIND_STALE_BINDING_RECOVERY,
+    }
+)
 
 
 _SAFE_SEGMENT_RE = re.compile(r"[^A-Za-z0-9._+-]+")
@@ -114,6 +150,7 @@ def state_key(
     org: str | None = None,
     repo: str | None = None,
     profile_identity: str | None = None,
+    instance_id: str | None = None,
 ) -> str:
     """Build durable filename key.
 
@@ -121,17 +158,20 @@ def state_key(
     so the primary key is kind + profile identity. Remote/org/repo are stored
     inside the payload and validated on load (#559), which lets a later daemon
     process recover state without already knowing the remote argument.
+
+    *instance_id* extends the key for kinds where one-per-profile collides —
+    concurrent same-profile sessions each own their reviewer-lease shadow
+    (#702 F5), keyed by their lease session id so a later heartbeat can never
+    overwrite or misattribute another session's crash evidence.
     """
     # Keep remote/org/repo parameters for API stability / future kinds; they are
     # intentionally not part of the filename for session-scoped proofs.
     _ = (remote, org, repo)
-    return "-".join(
+    parts = [kind, profile_identity or "unknown-profile"]
-        _sanitize_segment(part)
+    instance = (instance_id or "").strip()
-        for part in (
+    if instance:
-            kind,
+        parts.append(instance)
-            profile_identity or "unknown-profile",
+    return "-".join(_sanitize_segment(part) for part in parts)
-        )
-    )
 
 
 def state_file_path(
@@ -142,6 +182,7 @@ def state_file_path(
     repo: str | None = None,
     profile_identity: str | None = None,
     state_dir: str | None = None,
+    instance_id: str | None = None,
 ) -> str:
     root = (state_dir or default_state_dir()).strip()
     name = state_key(
@@ -150,6 +191,7 @@ def state_file_path(
         org=org,
         repo=repo,
         profile_identity=profile_identity,
+        instance_id=instance_id,
     )
     return os.path.join(root, f"{name}.json")
 
@@ -228,6 +270,16 @@ def _write_json(path: str, data: dict[str, Any]) -> None:
                 pass
 
 
+def is_recovery_critical_record(record: dict[str, Any] | None, kind: str | None = None) -> bool:
+    """True when a durable record must outlive the generic session-state TTL."""
+    if not record and not kind:
+        return False
+    record_kind = ((record or {}).get("kind") or kind or "").strip()
+    if record_kind in RECOVERY_CRITICAL_KINDS:
+        return True
+    return bool((record or {}).get("recovery_critical"))
+
+
 def identity_match_reasons(
     record: dict[str, Any] | None,
     *,
@@ -270,7 +322,9 @@ def identity_match_reasons(
         reasons.append("session state missing recorded_at timestamp (fail closed)")
     else:
         age = _now_utc() - recorded_at
-        if age > timedelta(hours=ttl_hours()):
+        kind = (record.get("kind") or "").strip()
+        ttl_exempt = is_recovery_critical_record(record, kind=kind)
+        if age > timedelta(hours=ttl_hours()) and not ttl_exempt:
             reasons.append(
                 f"session state expired after {ttl_hours():g}h (fail closed)"
             )
@@ -279,6 +333,113 @@ def identity_match_reasons(
     return reasons
 
 
+def inspect_state_envelope(
+    *,
+    kind: str,
+    remote: str | None = None,
+    org: str | None = None,
+    repo: str | None = None,
+    profile_identity: str | None = None,
+    state_dir: str | None = None,
+) -> dict[str, Any]:
+    """Read-only disk inspection for assessment when TTL would otherwise hide state (#720).
+
+    Does **not** apply identity/TTL rejection to the returned presence flags.
+    Callers use this to distinguish:
+      * no file on disk
+      * file present but TTL would reject a non-critical kind
+      * recovery-critical ledger (e.g. KIND_DECISION_LOCK) still loadable
+    Never mutates files. Never returns secrets.
+    """
+    profile = current_profile_identity(profile_identity=profile_identity)
+    path = state_file_path(
+        kind=kind,
+        remote=remote,
+        org=org,
+        repo=repo,
+        profile_identity=profile,
+        state_dir=state_dir,
+    )
+    result: dict[str, Any] = {
+        "kind": kind,
+        "profile_identity": profile,
+        "path_basename": os.path.basename(path),
+        "on_disk": False,
+        "has_payload": False,
+        "recorded_at": None,
+        "updated_at": None,
+        "age_hours": None,
+        "ttl_hours": ttl_hours(),
+        "age_exceeds_default_ttl": False,
+        "recovery_critical": kind in RECOVERY_CRITICAL_KINDS,
+        "ttl_exempt": kind in RECOVERY_CRITICAL_KINDS,
+        "would_ttl_reject": False,
+        "identity_reasons": [],
+        "summary": "no session-state file on disk",
+    }
+    if not path or not os.path.exists(path):
+        return result
+    result["on_disk"] = True
+    envelope = _read_json(path)
+    if not envelope:
+        result["summary"] = "session-state file present but unreadable or empty"
+        return result
+    payload = envelope.get("payload")
+    merged: dict[str, Any] = dict(payload) if isinstance(payload, dict) else {}
+    result["has_payload"] = isinstance(payload, dict)
+    for key in (
+        "kind",
+        "remote",
+        "org",
+        "repo",
+        "profile_identity",
+        "session_profile_lock",
+        "recorded_at",
+        "updated_at",
+        "writer_pid",
+        "recovery_critical",
+    ):
+        if key in envelope and key not in merged:
+            merged[key] = envelope[key]
+    if not merged.get("kind"):
+        merged["kind"] = kind
+    recorded_at = _parse_iso(merged.get("recorded_at") or merged.get("updated_at"))
+    result["recorded_at"] = merged.get("recorded_at") or merged.get("updated_at")
+    result["updated_at"] = merged.get("updated_at") or merged.get("recorded_at")
+    if recorded_at is not None:
+        age = _now_utc() - recorded_at
+        age_hours = age.total_seconds() / 3600.0
+        result["age_hours"] = age_hours
+        result["age_exceeds_default_ttl"] = age > timedelta(hours=ttl_hours())
+    ttl_exempt = is_recovery_critical_record(merged, kind=kind)
+    result["recovery_critical"] = ttl_exempt
+    result["ttl_exempt"] = ttl_exempt
+    identity_reasons = identity_match_reasons(
+        merged,
+        remote=remote,
+        org=org,
+        repo=repo,
+        profile_identity=profile,
+    )
+    result["identity_reasons"] = list(identity_reasons)
+    result["would_ttl_reject"] = any("expired" in r for r in identity_reasons)
+    if result["has_payload"] and ttl_exempt:
+        result["summary"] = (
+            "recovery-critical session-state present on disk and TTL-exempt; "
+            "load via load_state for full payload"
+        )
+    elif result["has_payload"] and result["would_ttl_reject"]:
+        result["summary"] = (
+            "session-state file present on disk but generic TTL would reject load "
+            f"(age_hours={result.get('age_hours')!r}, ttl={ttl_hours():g}h)"
+        )
+    elif result["has_payload"]:
+        result["summary"] = "session-state file present and within TTL / identity gates"
+    else:
+        result["summary"] = "session-state file present without a dict payload"
+    return result
+
+
 def load_state(
     *,
     kind: str,
@@ -287,6 +448,7 @@ def load_state(
     repo: str | None = None,
     profile_identity: str | None = None,
     state_dir: str | None = None,
+    instance_id: str | None = None,
 ) -> dict[str, Any] | None:
     """Load durable state payload when identity checks pass."""
     profile = current_profile_identity(profile_identity=profile_identity)
@@ -297,6 +459,7 @@ def load_state(
         repo=repo,
         profile_identity=profile,
         state_dir=state_dir,
+        instance_id=instance_id,
     )
     lock_path = f"{path}.lock"
     with _exclusive_file_lock(lock_path):
@@ -342,6 +505,7 @@ def save_state(
     repo: str | None = None,
     profile_identity: str | None = None,
     state_dir: str | None = None,
+    instance_id: str | None = None,
 ) -> dict[str, Any] | None:
     """Persist or clear durable state for the given session identity key."""
     profile = current_profile_identity(
@@ -354,6 +518,11 @@ def save_state(
     key_remote = remote if remote is not None else (payload or {}).get("remote")
     key_org = org if org is not None else (payload or {}).get("org")
     key_repo = repo if repo is not None else (payload or {}).get("repo")
+    key_instance = (
+        (instance_id or "").strip()
+        or str((payload or {}).get("instance_id") or "").strip()
+        or None
+    )
 
     root = _ensure_state_dir(state_dir)
     path = state_file_path(
@@ -363,6 +532,7 @@ def save_state(
         repo=key_repo,
         profile_identity=profile,
         state_dir=root,
+        instance_id=key_instance,
     )
     lock_path = f"{path}.lock"
     with _exclusive_file_lock(lock_path):
@@ -392,6 +562,8 @@ def save_state(
             body["repo"] = key_repo
         # Stamp session-state authority used for this write (#695 AC2 / AC6).
         body.setdefault("session_state_dir", root)
+        if key_instance:
+            body["instance_id"] = key_instance
         try:
             import mcp_daemon_guard
 
@@ -425,6 +597,8 @@ def save_state(
             "transport": body.get("transport"),
             "payload": body,
         }
+        if key_instance:
+            envelope["instance_id"] = key_instance
         _write_json(path, envelope)
         return dict(body)
 
@@ -437,6 +611,7 @@ def clear_state(
     repo: str | None = None,
     profile_identity: str | None = None,
     state_dir: str | None = None,
+    instance_id: str | None = None,
 ) -> None:
     save_state(
         kind=kind,
@@ -446,4 +621,229 @@ def clear_state(
         repo=repo,
         profile_identity=profile_identity,
         state_dir=state_dir,
+        instance_id=instance_id,
     )
+
+def list_states(
+    *,
+    kind: str,
+    profile_identity: str | None = None,
+    state_dir: str | None = None,
+) -> list[dict[str, Any]]:
+    """Enumerate durable records of *kind* across all instance keys.
+
+    Crash recovery cannot know which session id left a shadow behind, so it
+    must be able to enumerate every record of a kind (#702 F5). Identity
+    checks (profile / TTL policy) apply per record exactly as in
+    :func:`load_state`; records that fail closed are omitted.
+    """
+    root = (state_dir or default_state_dir()).strip()
+    if not root or not os.path.isdir(root):
+        return []
+    profile = current_profile_identity(profile_identity=profile_identity)
+    results: list[dict[str, Any]] = []
+    try:
+        names = sorted(os.listdir(root))
+    except OSError:
+        return []
+    for name in names:
+        if not name.endswith(".json") or name.startswith("."):
+            continue
+        envelope = _read_json(os.path.join(root, name))
+        if not envelope or envelope.get("kind") != kind:
+            continue
+        payload = envelope.get("payload")
+        if not isinstance(payload, dict):
+            continue
+        merged = dict(payload)
+        for key in (
+            "kind",
+            "remote",
+            "org",
+            "repo",
+            "profile_identity",
+            "session_profile_lock",
+            "recorded_at",
+            "updated_at",
+            "writer_pid",
+        ):
+            if key in envelope and key not in merged:
+                merged[key] = envelope[key]
+        if identity_match_reasons(merged, profile_identity=profile):
+            continue
+        results.append(merged)
+    return results
+
+def list_decision_lock_profile_identities(
+    state_dir: str | None = None,
+) -> list[str]:
+    """Return profile identities that have a durable review_decision_lock file (#709).
+
+    Filename form: ``review_decision_lock-<profile>.json`` (see ``state_key``).
+    Does not validate TTL or identity — callers must load via ``load_state``.
+    """
+    root = (state_dir or default_state_dir()).strip()
+    if not root or not os.path.isdir(root):
+        return []
+    prefix = f"{_sanitize_segment(KIND_DECISION_LOCK)}-"
+    suffix = ".json"
+    found: list[str] = []
+    try:
+        names = os.listdir(root)
+    except OSError:
+        return []
+    for name in names:
+        if not name.startswith(prefix) or not name.endswith(suffix):
+            continue
+        if name.endswith(".lock"):
+            continue
+        mid = name[len(prefix) : -len(suffix)]
+        if mid:
+            found.append(mid)
+    return sorted(set(found))
+
+
+def load_state_for_profile(
+    *,
+    kind: str,
+    profile_identity: str,
+    remote: str | None = None,
+    org: str | None = None,
+    repo: str | None = None,
+    state_dir: str | None = None,
+    skip_identity_match: bool = False,
+    enforce_repo_scope: bool = True,
+) -> dict[str, Any] | None:
+    """Load durable state for an explicit profile identity (#709 cross-profile).
+
+    When *skip_identity_match* is True, still requires the file's recorded
+    profile_identity to equal the requested profile (anti-stomp), but does not
+    require the *active* session identity to match — needed so a merger can
+    inspect a reviewer lock after merge.
+
+    #709 F3: remote/org/repo filter reasons are **enforced** (not merely
+    computed). When *enforce_repo_scope* is True (default) and the caller
+    supplies remote/org/repo, mismatches fail closed with ``None``.
+    """
+    # #709 F3: refuse traversal / malformed profile identities.
+    try:
+        from irrecoverable_provenance import assess_profile_path_identity
+
+        path_gate = assess_profile_path_identity(profile_identity)
+        if not path_gate.get("valid"):
+            return None
+    except Exception:
+        # Module may be mid-import in edge bootstraps; fall through to
+        # conservative checks below.
+        raw = str(profile_identity or "")
+        if ".." in raw or "/" in raw or "\\" in raw or "\x00" in raw:
+            return None
+
+    profile = current_profile_identity(profile_identity=profile_identity)
+    root = _ensure_state_dir(state_dir)
+    # Refuse symlink escape of the state root (#709 F3).
+    try:
+        real_root = os.path.realpath(root)
+        path = state_file_path(
+            kind=kind,
+            remote=remote,
+            org=org,
+            repo=repo,
+            profile_identity=profile,
+            state_dir=root,
+        )
+        real_path = os.path.realpath(path) if os.path.exists(path) else path
+        if os.path.exists(path) and not str(real_path).startswith(
+            str(real_root) + os.sep
+        ) and str(real_path) != str(real_root):
+            return None
+    except OSError:
+        return None
+
+    path = state_file_path(
+        kind=kind,
+        remote=remote,
+        org=org,
+        repo=repo,
+        profile_identity=profile,
+        state_dir=root,
+    )
+    lock_path = f"{path}.lock"
+    with _exclusive_file_lock(lock_path):
+        envelope = _read_json(path)
+    if not envelope:
+        return None
+    payload = envelope.get("payload")
+    if not isinstance(payload, dict):
+        return None
+    merged = dict(payload)
+    for key in (
+        "kind",
+        "remote",
+        "org",
+        "repo",
+        "profile_identity",
+        "session_profile_lock",
+        "recorded_at",
+        "updated_at",
+        "writer_pid",
+    ):
+        if key in envelope and key not in merged:
+            merged[key] = envelope[key]
+    stored = (merged.get("profile_identity") or "").strip()
+    if stored and stored != profile:
+        return None
+    if skip_identity_match:
+        # Enforce remote/org/repo when caller provides them (#709 F3).
+        # Drop only *active session* profile-identity mismatches; keep
+        # expiry, spoof, and repository-scope reasons.
+        scope_remote = remote if enforce_repo_scope else None
+        scope_org = org if enforce_repo_scope else None
+        scope_repo = repo if enforce_repo_scope else None
+        reasons = identity_match_reasons(
+            merged,
+            remote=scope_remote,
+            org=scope_org,
+            repo=scope_repo,
+            profile_identity=stored or profile,
+        )
+        filtered = [
+            r
+            for r in reasons
+            if "profile identity mismatch" not in r
+            or (stored and stored != profile)
+        ]
+        # When caller requested a specific remote/org/repo, also fail if the
+        # stored record lacks those identity fields entirely (legacy incomplete).
+        if enforce_repo_scope:
+            for field, want in (
+                ("remote", remote),
+                ("org", org),
+                ("repo", repo),
+            ):
+                want_s = (want or "").strip()
+                if not want_s:
+                    continue
+                have = (str(merged.get(field) or "")).strip()
+                if not have:
+                    filtered.append(
+                        f"session state {field} missing on durable record "
+                        f"(expected={want_s!r}; fail closed, #709 F3)"
+                    )
+                elif have != want_s:
+                    # identity_match_reasons already adds mismatch; ensure kept
+                    pass
+        if filtered:
+            return None
+        return merged
+    reasons = identity_match_reasons(
+        merged,
+        remote=remote,
+        org=org,
+        repo=repo,
+        profile_identity=profile,
+    )
+    if reasons:
+        return None
+    return merged
+

mcp_tool_error_boundary.py

@@ -0,0 +1,451 @@
+"""MCP tool-boundary error mapping for known Gitea client failures (#699).
+
+Known authentication / authorization / network / configuration failures leave
+the tool boundary as a sanitized structured ``CallToolResult`` with
+``isError=True``. Stdio transport remains connected.
+
+Design constraints (reviewer-ratified, #699 / PR #701):
+
+1. **No secret material** in tool results or daemon logs. Messages are fixed
+   constants keyed by ``reason_code``; HTTP bodies / exception text never
+   surface. Sanitization failure fails closed to ``internal_error``.
+2. **Narrow boundary** wraps the original FastMCP ``Tool.run`` success path;
+   re-raises framework control-flow exceptions (``UrlElicitationRequiredError``);
+   maps only typed client failures. Installation is idempotent.
+3. **No RuntimeError substring heuristics.** Only explicit typed
+   authentication / authorization / network / configuration exceptions
+   receive those labels.
+4. **HTTP classification** lives in ``gitea_auth.classify_http_status``;
+   every 403 is authorization-class.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import sys
+from typing import Any
+
+logger = logging.getLogger("gitea_mcp.tool_error_boundary")
+
+# Stable reason codes (#699).
+REASON_AUTH_FAILED = "auth_failed"
+REASON_AUTH_INVALID_TOKEN = "auth_invalid_token"
+REASON_AUTHZ_INSUFFICIENT_SCOPE = "authz_insufficient_scope"
+REASON_AUTHZ_DENIED = "authz_denied"
+REASON_NETWORK_ERROR = "network_error"
+REASON_CONFIG_ERROR = "config_error"
+REASON_INTERNAL_ERROR = "internal_error"
+REASON_UPSTREAM_UNAVAILABLE = "upstream_unavailable"
+REASON_HTTP_ERROR = "http_error"
+
+ERROR_CLASS_AUTHENTICATION = "authentication"
+ERROR_CLASS_AUTHORIZATION = "authorization"
+ERROR_CLASS_NETWORK = "network"
+ERROR_CLASS_CONFIGURATION = "configuration"
+ERROR_CLASS_INTERNAL = "internal"
+ERROR_CLASS_UPSTREAM = "upstream"
+
+# Fixed, secret-free operator messages. Never interpolate HTTP bodies,
+# Keychain contents, tokens, or arbitrary exception text.
+FIXED_MESSAGES: dict[str, str] = {
+    REASON_AUTH_FAILED: "Gitea authentication failed",
+    REASON_AUTH_INVALID_TOKEN: (
+        "Gitea authentication failed: invalid or revoked credentials"
+    ),
+    REASON_AUTHZ_INSUFFICIENT_SCOPE: (
+        "Gitea authorization failed: insufficient token scope"
+    ),
+    REASON_AUTHZ_DENIED: "Gitea authorization failed: access denied",
+    REASON_NETWORK_ERROR: "Network error contacting Gitea",
+    REASON_CONFIG_ERROR: "Gitea configuration or credential resolution failed",
+    REASON_INTERNAL_ERROR: "Internal tool error",
+    REASON_UPSTREAM_UNAVAILABLE: "Gitea upstream unavailable",
+    REASON_HTTP_ERROR: "Gitea HTTP request failed",
+}
+
+_INSTALL_FLAG = "_gitea_auth_boundary_installed"
+_ORIGINAL_ATTR = "_gitea_auth_boundary_original"
+
+
+def fixed_message(reason_code: str) -> str:
+    """Return the fixed sanitized message for *reason_code* (fail closed)."""
+    return FIXED_MESSAGES.get(reason_code, FIXED_MESSAGES[REASON_INTERNAL_ERROR])
+
+
+def _safe_profile_name() -> str | None:
+    try:
+        from gitea_auth import get_profile
+
+        name = (get_profile() or {}).get("profile_name")
+        if name is None:
+            return None
+        text = str(name).strip()
+        # Profile names are non-secret identifiers; still bound length.
+        return text[:80] if text else None
+    except Exception:
+        return None
+
+
+def _is_framework_control_flow(exc: BaseException) -> bool:
+    """True for exceptions the MCP framework must re-raise unchanged."""
+    try:
+        from mcp.shared.exceptions import UrlElicitationRequiredError
+
+        if isinstance(exc, UrlElicitationRequiredError):
+            return True
+    except Exception:
+        pass
+    # BaseException subclasses that must never become tool isError payloads.
+    if isinstance(exc, (KeyboardInterrupt, SystemExit, GeneratorExit)):
+        return True
+    return False
+
+
+def is_known_client_failure(exc: BaseException) -> bool:
+    """True only for explicit typed Gitea client / config failures.
+
+    Never true for arbitrary ``RuntimeError`` (reviewer finding #3).
+    """
+    try:
+        import gitea_auth
+
+        if isinstance(
+            exc,
+            (
+                gitea_auth.GiteaAuthError,
+                gitea_auth.GiteaAuthzError,
+                gitea_auth.GiteaNetworkError,
+                gitea_auth.GiteaConfigError,
+                gitea_auth.GiteaHttpError,
+            ),
+        ):
+            return True
+    except Exception:
+        return False
+    try:
+        import gitea_config
+
+        if isinstance(exc, gitea_config.ConfigError):
+            return True
+    except Exception:
+        pass
+    return False
+
+
+def classify_exception(exc: BaseException) -> dict[str, Any]:
+    """Return structured classification with **fixed** messages only.
+
+    Only typed authentication / authorization / network / configuration
+    failures receive those labels. Unexpected programming failures are
+    ``internal_error``. HTTP bodies and exception text are never copied
+    into ``message``.
+    """
+    try:
+        return _classify_exception_impl(exc)
+    except Exception:
+        # Fail closed: sanitization / classification failure must not leak.
+        return {
+            "reason_code": REASON_INTERNAL_ERROR,
+            "error_class": ERROR_CLASS_INTERNAL,
+            "http_status": None,
+            "message": fixed_message(REASON_INTERNAL_ERROR),
+            "transport_survives": True,
+        }
+
+
+def _classify_exception_impl(exc: BaseException) -> dict[str, Any]:
+    import gitea_auth
+
+    if isinstance(exc, gitea_auth.GiteaAuthError):
+        code = getattr(exc, "reason_code", None) or REASON_AUTH_INVALID_TOKEN
+        if code not in (
+            REASON_AUTH_FAILED,
+            REASON_AUTH_INVALID_TOKEN,
+        ):
+            code = REASON_AUTH_INVALID_TOKEN
+        return {
+            "reason_code": code,
+            "error_class": ERROR_CLASS_AUTHENTICATION,
+            "http_status": getattr(exc, "http_status", None) or 401,
+            "message": fixed_message(code),
+            "transport_survives": True,
+        }
+    if isinstance(exc, gitea_auth.GiteaAuthzError):
+        code = getattr(exc, "reason_code", None) or REASON_AUTHZ_DENIED
+        if code not in (REASON_AUTHZ_DENIED, REASON_AUTHZ_INSUFFICIENT_SCOPE):
+            code = REASON_AUTHZ_DENIED
+        return {
+            "reason_code": code,
+            "error_class": ERROR_CLASS_AUTHORIZATION,
+            "http_status": getattr(exc, "http_status", None) or 403,
+            "message": fixed_message(code),
+            "transport_survives": True,
+        }
+    if isinstance(exc, gitea_auth.GiteaNetworkError):
+        return {
+            "reason_code": REASON_NETWORK_ERROR,
+            "error_class": ERROR_CLASS_NETWORK,
+            "http_status": getattr(exc, "http_status", None),
+            "message": fixed_message(REASON_NETWORK_ERROR),
+            "transport_survives": True,
+        }
+    if isinstance(exc, gitea_auth.GiteaConfigError):
+        return {
+            "reason_code": REASON_CONFIG_ERROR,
+            "error_class": ERROR_CLASS_CONFIGURATION,
+            "http_status": getattr(exc, "http_status", None),
+            "message": fixed_message(REASON_CONFIG_ERROR),
+            "transport_survives": True,
+        }
+    if isinstance(exc, gitea_auth.GiteaHttpError):
+        code = getattr(exc, "reason_code", None) or REASON_HTTP_ERROR
+        if code == REASON_UPSTREAM_UNAVAILABLE:
+            error_class = ERROR_CLASS_UPSTREAM
+        else:
+            error_class = ERROR_CLASS_INTERNAL
+            code = REASON_HTTP_ERROR
+        return {
+            "reason_code": code,
+            "error_class": error_class,
+            "http_status": getattr(exc, "http_status", None),
+            "message": fixed_message(code),
+            "transport_survives": True,
+        }
+
+    try:
+        import gitea_config
+
+        if isinstance(exc, gitea_config.ConfigError):
+            return {
+                "reason_code": REASON_CONFIG_ERROR,
+                "error_class": ERROR_CLASS_CONFIGURATION,
+                "http_status": None,
+                "message": fixed_message(REASON_CONFIG_ERROR),
+                "transport_survives": True,
+            }
+    except Exception:
+        pass
+
+    # Unwrap FastMCP ToolError cause when the original was a typed failure.
+    cause = getattr(exc, "__cause__", None)
+    if cause is not None and cause is not exc and is_known_client_failure(cause):
+        return _classify_exception_impl(cause)
+
+    # No message-substring authentication heuristics (reviewer finding #3).
+    return {
+        "reason_code": REASON_INTERNAL_ERROR,
+        "error_class": ERROR_CLASS_INTERNAL,
+        "http_status": None,
+        "message": fixed_message(REASON_INTERNAL_ERROR),
+        "transport_survives": True,
+    }
+
+
+def build_structured_error_payload(
+    classification: dict[str, Any],
+    *,
+    tool_name: str | None = None,
+    profile_name: str | None = None,
+) -> dict[str, Any]:
+    """LLM-safe structured payload — fixed message + typed metadata only."""
+    try:
+        reason = str(classification.get("reason_code") or REASON_INTERNAL_ERROR)
+        message = fixed_message(reason)
+        # Refuse to emit any classification message that is not the fixed constant.
+        if classification.get("message") != message:
+            message = fixed_message(reason)
+        payload: dict[str, Any] = {
+            "success": False,
+            "isError": True,
+            "reason_code": reason if reason in FIXED_MESSAGES else REASON_INTERNAL_ERROR,
+            "error_class": classification.get("error_class") or ERROR_CLASS_INTERNAL,
+            "message": message,
+            "transport_survives": True,
+            "retryable": classification.get("error_class")
+            in {
+                ERROR_CLASS_AUTHENTICATION,
+                ERROR_CLASS_NETWORK,
+                ERROR_CLASS_CONFIGURATION,
+            },
+        }
+        status = classification.get("http_status")
+        if isinstance(status, int):
+            payload["http_status"] = status
+        if tool_name and isinstance(tool_name, str):
+            # Tool names are identifiers, not secrets; bound length.
+            payload["tool"] = tool_name[:120]
+        if profile_name and isinstance(profile_name, str):
+            payload["profile"] = profile_name[:80]
+        return payload
+    except Exception:
+        return {
+            "success": False,
+            "isError": True,
+            "reason_code": REASON_INTERNAL_ERROR,
+            "error_class": ERROR_CLASS_INTERNAL,
+            "message": fixed_message(REASON_INTERNAL_ERROR),
+            "transport_survives": True,
+            "retryable": False,
+        }
+
+
+def log_sanitized_daemon_reason(
+    classification: dict[str, Any],
+    *,
+    tool_name: str | None = None,
+    stream=None,
+) -> None:
+    """Daemon log: reason codes only — never exception text or response bodies."""
+    stream = stream if stream is not None else sys.stderr
+    try:
+        reason = classification.get("reason_code") or REASON_INTERNAL_ERROR
+        error_class = classification.get("error_class") or ERROR_CLASS_INTERNAL
+        # Only emit known tokens; never classification['message'] from callers
+        # that might have been poisoned.
+        if reason not in FIXED_MESSAGES:
+            reason = REASON_INTERNAL_ERROR
+            error_class = ERROR_CLASS_INTERNAL
+        parts = [
+            "mcp_tool_error",
+            f"reason_code={reason}",
+            f"error_class={error_class}",
+        ]
+        if tool_name and isinstance(tool_name, str):
+            parts.append(f"tool={tool_name[:120]}")
+        status = classification.get("http_status")
+        if isinstance(status, int):
+            parts.append(f"http_status={status}")
+        # Intentionally no detail= / message= field — secrets lived there.
+        line = " ".join(parts)
+        stream.write(line + "\n")
+        if hasattr(stream, "flush"):
+            stream.flush()
+        logger.warning(line)
+    except Exception:
+        # Fail closed: never fall back to logging the exception.
+        try:
+            stream.write(
+                "mcp_tool_error reason_code=internal_error "
+                "error_class=internal\n"
+            )
+            if hasattr(stream, "flush"):
+                stream.flush()
+        except Exception:
+            pass
+
+
+def to_call_tool_result(
+    exc: BaseException,
+    *,
+    tool_name: str | None = None,
+    profile_name: str | None = None,
+    log: bool = True,
+) -> Any:
+    """Build a FastMCP ``CallToolResult`` with ``isError=True`` for *exc*."""
+    from mcp.types import CallToolResult, TextContent
+
+    try:
+        classification = classify_exception(exc)
+        if log:
+            log_sanitized_daemon_reason(classification, tool_name=tool_name)
+        payload = build_structured_error_payload(
+            classification, tool_name=tool_name, profile_name=profile_name
+        )
+        text = json.dumps(payload, indent=2, sort_keys=True)
+        return CallToolResult(
+            content=[TextContent(type="text", text=text)],
+            structuredContent=payload,
+            isError=True,
+        )
+    except Exception:
+        # Absolute fail-closed path — no exception text.
+        fallback = {
+            "success": False,
+            "isError": True,
+            "reason_code": REASON_INTERNAL_ERROR,
+            "error_class": ERROR_CLASS_INTERNAL,
+            "message": fixed_message(REASON_INTERNAL_ERROR),
+            "transport_survives": True,
+            "retryable": False,
+        }
+        return CallToolResult(
+            content=[
+                TextContent(
+                    type="text",
+                    text=json.dumps(fallback, indent=2, sort_keys=True),
+                )
+            ],
+            structuredContent=fallback,
+            isError=True,
+        )
+
+
+def install_tool_run_boundary(Tool) -> bool:
+    """Install a **narrow** error boundary around FastMCP ``Tool.run``.
+
+    Strategy (preserves framework semantics):
+
+    * Call the **original** ``Tool.run`` for the success path (async, return
+      types, convert_result, protocol behavior unchanged).
+    * Re-raise ``UrlElicitationRequiredError`` and other control-flow
+      exceptions without mapping to ``internal_error``.
+    * Map typed Gitea client failures (and ToolError whose ``__cause__`` is
+      typed) to structured ``CallToolResult(isError=True)``.
+    * Map remaining unexpected tool failures to fixed ``internal_error``
+      isError results (transport survival) without secret-bearing text.
+    * Idempotent: second install is a no-op and returns ``False``.
+    """
+    if getattr(Tool.run, _INSTALL_FLAG, False):
+        return False
+
+    from mcp.server.fastmcp.exceptions import ToolError
+    from mcp.shared.exceptions import UrlElicitationRequiredError
+
+    original_run = Tool.run
+
+    async def run_boundary(
+        self,
+        arguments: dict[str, Any],
+        context=None,
+        convert_result: bool = False,
+    ) -> Any:
+        try:
+            return await original_run(
+                self,
+                arguments,
+                context=context,
+                convert_result=convert_result,
+            )
+        except UrlElicitationRequiredError:
+            # Framework control-flow — must not become internal_error.
+            raise
+        except BaseException as exc:
+            if _is_framework_control_flow(exc):
+                raise
+            if not isinstance(exc, Exception):
+                raise
+
+            # Prefer typed cause under FastMCP ToolError wrappers.
+            target: BaseException = exc
+            if isinstance(exc, ToolError) and exc.__cause__ is not None:
+                target = exc.__cause__
+
+            # Known client failures → structured isError with reason codes.
+            # Unexpected failures → fixed internal_error isError (no secrets).
+            profile_name = _safe_profile_name()
+            return to_call_tool_result(
+                target,
+                tool_name=getattr(self, "name", None),
+                profile_name=profile_name,
+            )
+
+    setattr(run_boundary, _INSTALL_FLAG, True)
+    setattr(run_boundary, _ORIGINAL_ATTR, original_run)
+    Tool.run = run_boundary  # type: ignore[method-assign]
+    return True
+
+
+def boundary_is_installed(Tool) -> bool:
+    """True when the #699 boundary is active on *Tool.run*."""
+    return bool(getattr(Tool.run, _INSTALL_FLAG, False))

migrate_profiles.py

@@ -20,12 +20,114 @@ if PROJECT_ROOT not in sys.path:
 import gitea_config
 
 
-AUTHOR_DEFAULT_ALLOWED = ["read", "branch", "commit", "push", "open_pr", "comment"]
+# Defaults emit *canonical* operation names only. Shorthand that is not in
-AUTHOR_DEFAULT_FORBIDDEN = ["approve", "request_changes", "merge"]
+# gitea_config.GITEA_OPERATION_ALIASES (e.g. ``pr.close``, ``issue.close``)
-REVIEWER_DEFAULT_ALLOWED = [
+# is silently dropped by the production loader and must never appear here.
-    "read", "review", "comment", "approve", "request_changes", "merge"
+AUTHOR_DEFAULT_ALLOWED = [
+    "gitea.read",
+    "gitea.branch.create",
+    "gitea.repo.commit",
+    "gitea.branch.push",
+    "gitea.pr.create",
+    "gitea.pr.comment",
 ]
-REVIEWER_DEFAULT_FORBIDDEN = ["branch", "commit", "push", "open_pr"]
+AUTHOR_DEFAULT_FORBIDDEN = [
+    "gitea.pr.approve",
+    "gitea.pr.request_changes",
+    "gitea.pr.merge",
+]
+REVIEWER_DEFAULT_ALLOWED = [
+    "gitea.read",
+    "gitea.pr.review",
+    "gitea.pr.comment",
+    "gitea.pr.approve",
+    "gitea.pr.request_changes",
+    "gitea.pr.merge",
+]
+REVIEWER_DEFAULT_FORBIDDEN = [
+    "gitea.branch.create",
+    "gitea.repo.commit",
+    "gitea.branch.push",
+    "gitea.pr.create",
+]
+# Required reconciler ops (read + pr.close) plus recommended comment/close and
+# branch.delete for guarded merged-PR cleanup. All names must normalize via
+# gitea_config.normalize_operation without being dropped.
+RECONCILER_DEFAULT_ALLOWED = [
+    "gitea.read",
+    "gitea.pr.close",
+    "gitea.pr.comment",
+    "gitea.issue.comment",
+    "gitea.issue.close",
+    "gitea.branch.delete",
+]
+RECONCILER_DEFAULT_FORBIDDEN = [
+    "gitea.pr.approve",
+    "gitea.pr.merge",
+    "gitea.pr.review",
+    "gitea.pr.create",
+    "gitea.branch.push",
+    "gitea.repo.commit",
+]
+
+# Migration-only expansions for common shorthands that are *not* in
+# GITEA_OPERATION_ALIASES. Emitted output is always the canonical form so a
+# second canonicalize pass is a no-op (idempotent).
+_MIGRATION_ONLY_ALIASES = {
+    "pr.close": "gitea.pr.close",
+    "pr.comment": "gitea.pr.comment",
+    "issue.close": "gitea.issue.close",
+    "branch.delete": "gitea.branch.delete",
+}
+
+# Reconciler required ops that must survive migration (from reconciler_profile).
+RECONCILER_REQUIRED_CANONICAL = ("gitea.read", "gitea.pr.close")
+
+
+def canonicalize_operation(op: str) -> str:
+    """Return a canonical operation name accepted by the production loader.
+
+    Fail closed on unknown/ambiguous spellings so required permissions cannot
+    be silently dropped by ``check_operation`` later.
+    """
+    if not isinstance(op, str) or not op.strip():
+        raise ValueError("operation must be a non-empty string (fail closed)")
+    op = op.strip()
+    try:
+        return gitea_config.normalize_operation(op)
+    except gitea_config.ConfigError:
+        pass
+    if op in _MIGRATION_ONLY_ALIASES:
+        return _MIGRATION_ONLY_ALIASES[op]
+    raise ValueError(
+        f"operation {op!r} cannot be canonicalized for migration "
+        "(unknown/ambiguous; fail closed — production loader would drop it)"
+    )
+
+
+def canonicalize_operations(ops, *, context: str = "operations") -> list[str]:
+    """Canonicalize a list of operations; preserve order, drop duplicates."""
+    if not isinstance(ops, list):
+        raise ValueError(f"{context} must be a list (fail closed)")
+    out: list[str] = []
+    seen: set[str] = set()
+    for entry in ops:
+        canon = canonicalize_operation(entry)
+        if canon not in seen:
+            seen.add(canon)
+            out.append(canon)
+    return out
+
+
+def _assert_reconciler_required_survive(allowed: list[str], profile_name: str) -> None:
+    """Fail visibly when migration would leave a reconciler without required ops."""
+    missing = [op for op in RECONCILER_REQUIRED_CANONICAL if op not in set(allowed)]
+    if missing:
+        raise ValueError(
+            f"Profile '{profile_name}' (reconciler) is missing required "
+            f"operation(s) after migration: {missing}. Refusing to emit a "
+            "profile that would silently fail pr.close / read (fail closed)."
+        )
 
 
 def infer_role(name, execution_profile):
@@ -90,9 +192,11 @@ def migrate_v1_to_v2(v1_data):
             ident_name = "reviewer"
         elif role == "author":
             ident_name = "author"
+        elif role == "reconciler":
+            ident_name = "reconciler"
         else:
             role = prof.get("role")
-            if role not in (None, "author", "reviewer"):
+            if role not in (None, "author", "reviewer", "reconciler"):
                 raise ValueError(
                     f"Profile '{name}' has unsupported role {role!r}"
                 )
@@ -124,20 +228,35 @@ def migrate_v1_to_v2(v1_data):
                 raise ValueError(
                     f"Profile '{name}' operation fields must be lists"
                 )
-            identity_data["allowed_operations"] = list(allowed)
+            try:
-            identity_data["forbidden_operations"] = list(forbidden)
+                identity_data["allowed_operations"] = canonicalize_operations(
+                    allowed, context=f"profile '{name}' allowed_operations"
+                )
+                identity_data["forbidden_operations"] = canonicalize_operations(
+                    forbidden, context=f"profile '{name}' forbidden_operations"
+                )
+            except ValueError as exc:
+                raise ValueError(f"Profile '{name}': {exc}") from exc
         elif role == "author":
             identity_data["allowed_operations"] = list(AUTHOR_DEFAULT_ALLOWED)
             identity_data["forbidden_operations"] = list(AUTHOR_DEFAULT_FORBIDDEN)
         elif role == "reviewer":
             identity_data["allowed_operations"] = list(REVIEWER_DEFAULT_ALLOWED)
             identity_data["forbidden_operations"] = list(REVIEWER_DEFAULT_FORBIDDEN)
+        elif role == "reconciler":
+            identity_data["allowed_operations"] = list(RECONCILER_DEFAULT_ALLOWED)
+            identity_data["forbidden_operations"] = list(RECONCILER_DEFAULT_FORBIDDEN)
         else:
             raise ValueError(
                 f"Profile '{name}' has no explicit operation lists and no "
                 "unambiguous author/reviewer role marker (fail closed)"
             )
 
+        if role == "reconciler":
+            _assert_reconciler_required_survive(
+                identity_data["allowed_operations"], name
+            )
+
         # Nest inside environments/services structure
         env = environments.setdefault(env_name, {})
         services = env.setdefault("services", {})

namespace_workspace_binding.py

@@ -56,23 +56,46 @@ def resolve_namespace_workspace(
     env: dict[str, str] | os._Environ | None = None,
     session_lease_worktree: str | None = None,
     profile_name: str | None = None,
+    demotions: list[str] | None = None,
+    verify_paths: bool = False,
 ) -> tuple[str, str]:
-    """Return ``(resolved_path, binding_source)`` for *role_kind*."""
+    """Return ``(resolved_path, binding_source)`` for *role_kind*.
+
+    With *verify_paths*, env-sourced candidates whose path no longer exists
+    are demoted (#702): a binding to a deleted worktree can never name a
+    valid task workspace, so resolution falls through to the next candidate.
+    Explicit arguments are never demoted — a caller-declared path must fail
+    loudly downstream rather than silently rebind. Demotion notes are
+    appended to *demotions* when provided. Runtime-context and mutation
+    guards resolve through :func:`resolve_namespace_mutation_context`, which
+    always verifies.
+    """
     env_map = env if env is not None else os.environ
     role = normalize_role_kind(role_kind, profile_name=profile_name)
     role_env_key = ROLE_WORKTREE_ENVS[role]
 
-    for candidate, source in (
+    for candidate, source, env_sourced in (
-        (worktree_path, "worktree_path argument"),
+        (worktree_path, "worktree_path argument", False),
-        (worktree, "worktree argument"),
+        (worktree, "worktree argument", False),
-        (_env_value(env_map, ACTIVE_WORKTREE_ENV), f"{ACTIVE_WORKTREE_ENV} environment variable"),
+        (_env_value(env_map, ACTIVE_WORKTREE_ENV),
-        (_env_value(env_map, role_env_key), f"{role_env_key} environment variable"),
+         f"{ACTIVE_WORKTREE_ENV} environment variable", True),
+        (_env_value(env_map, role_env_key),
+         f"{role_env_key} environment variable", True),
         (session_lease_worktree if role in {"reviewer", "merger"} else None,
-         "reviewer PR lease worktree"),
+         "reviewer PR lease worktree", False),
     ):
         text = (candidate or "").strip()
-        if text:
+        if not text:
-            return os.path.realpath(os.path.abspath(text)), source
+            continue
+        real = os.path.realpath(os.path.abspath(text))
+        if verify_paths and env_sourced and not os.path.isdir(real):
+            if demotions is not None:
+                demotions.append(
+                    f"{source} '{real}' demoted: path no longer exists "
+                    "(stale binding, #702)"
+                )
+            continue
+        return real, source
 
     return os.path.realpath(process_project_root), "MCP server process root (default)"
 
@@ -88,6 +111,7 @@ def resolve_namespace_mutation_context(
     profile_name: str | None = None,
 ) -> dict:
     """Shared workspace resolution for runtime_context and mutation guards."""
+    demotions: list[str] = []
     workspace, binding_source = resolve_namespace_workspace(
         role_kind=role_kind,
         worktree_path=worktree_path,
@@ -96,6 +120,8 @@ def resolve_namespace_mutation_context(
         env=env,
         session_lease_worktree=session_lease_worktree,
         profile_name=profile_name,
+        demotions=demotions,
+        verify_paths=True,
     )
     process_root = os.path.realpath(process_project_root)
     role = normalize_role_kind(role_kind, profile_name=profile_name)
@@ -111,7 +137,7 @@ def resolve_namespace_mutation_context(
         "workspace_path": workspace,
         "workspace_binding_source": binding_source,
         "workspace_role_kind": role,
-        "ignored_bindings": pollution.get("ignored_bindings") or [],
+        "ignored_bindings": demotions + (pollution.get("ignored_bindings") or []),
         "process_project_root": process_root,
         "canonical_repo_root": canonical_root,
         "roots_aligned": canonical_root == process_root,

reconciler_profile.py

@@ -18,6 +18,11 @@ RECONCILER_RECOMMENDED_OPERATIONS = (
     "gitea.pr.comment",
     "gitea.issue.comment",
     "gitea.issue.close",
+    # Merged-branch cleanup is reconciler-owned (task_capability_map maps
+    # cleanup_merged_pr_branch -> reconciler). The permission is only
+    # exercisable through the guarded gitea_cleanup_merged_pr_branch path
+    # (#514): merged proof, protected-branch refusal, explicit confirmation.
+    "gitea.branch.delete",
 )
 
 RECONCILER_FORBIDDEN_OPERATIONS = (

reviewer_pr_lease.py

@@ -407,18 +407,194 @@ def record_session_lease(
     if lease_provenance:
         stored["lease_provenance"] = dict(lease_provenance)
     _SESSION_LEASE = stored
+    _persist_session_lease_shadow(stored)
     return dict(_SESSION_LEASE)
 
 
 def clear_session_lease() -> None:
     global _SESSION_LEASE
+    prior = _SESSION_LEASE
     _SESSION_LEASE = None
+    _persist_session_lease_shadow(None, prior=prior)
 
 
 def get_session_lease() -> dict[str, Any] | None:
     return dict(_SESSION_LEASE) if _SESSION_LEASE else None
 
 
+def _persist_session_lease_shadow(
+    lease: dict[str, Any] | None,
+    *,
+    prior: dict[str, Any] | None = None,
+) -> None:
+    """Best-effort durable shadow of the in-session lease (#702).
+
+    A daemon that dies without teardown leaves this record behind, giving a
+    later process provable orphan evidence (owner pid + session id) for the
+    guarded cleanup path. Observability only — mutation gates never read it,
+    and failures here must never block lease operations.
+
+    Shadows are keyed by lease session id (#702 F5): concurrent same-profile
+    sessions each own a distinct record, so one session's heartbeat can never
+    overwrite or misattribute another's crash evidence. A sanctioned clear is
+    the terminal reconciliation of that record's lifecycle.
+    """
+    try:
+        import mcp_session_state as mss
+
+        if lease is None:
+            sid = ((prior or {}).get("session_id") or "").strip() or None
+            if sid:
+                mss.clear_state(
+                    kind=mss.KIND_REVIEWER_SESSION_LEASE, instance_id=sid
+                )
+            # Legacy single-slot record from pre-F5 code: clearing it is safe
+            # because collision-safe writes never target that key again.
+            mss.clear_state(kind=mss.KIND_REVIEWER_SESSION_LEASE)
+            return
+        mss.save_state(
+            kind=mss.KIND_REVIEWER_SESSION_LEASE,
+            instance_id=((lease.get("session_id") or "").strip() or None),
+            payload={
+                "pr_number": lease.get("pr_number"),
+                "session_id": lease.get("session_id"),
+                "comment_id": lease.get("comment_id"),
+                "candidate_head": lease.get("candidate_head"),
+                "worktree": lease.get("worktree"),
+                "phase": lease.get("phase"),
+                "reviewer_identity": lease.get("reviewer_identity"),
+                "profile": lease.get("profile"),
+                "lease_repo": lease.get("repo"),
+                "owner_pid": os.getpid(),
+            },
+        )
+    except Exception:
+        pass
+
+
+def load_session_lease_shadow(
+    session_id: str | None = None,
+) -> dict[str, Any] | None:
+    """Load the durable session-lease shadow left by this profile identity.
+
+    With *session_id*, load that session's collision-safe record (#702 F5),
+    falling back to the legacy single-slot record only when its payload names
+    the same session. Without *session_id*, return the legacy record or the
+    sole surviving instance record — ambiguity (multiple candidates) returns
+    ``None`` because a shadow that cannot be attributed proves nothing.
+    """
+    try:
+        import mcp_session_state as mss
+
+        sid = (session_id or "").strip()
+        if sid:
+            record = mss.load_state(
+                kind=mss.KIND_REVIEWER_SESSION_LEASE, instance_id=sid
+            )
+            if record is not None:
+                return record
+            legacy = mss.load_state(kind=mss.KIND_REVIEWER_SESSION_LEASE)
+            if legacy and (legacy.get("session_id") or "").strip() == sid:
+                return legacy
+            return None
+        legacy = mss.load_state(kind=mss.KIND_REVIEWER_SESSION_LEASE)
+        if legacy is not None:
+            return legacy
+        records = mss.list_states(kind=mss.KIND_REVIEWER_SESSION_LEASE)
+        if len(records) == 1:
+            return records[0]
+        return None
+    except Exception:
+        return None
+
+
+def _pid_alive(pid: int) -> bool:
+    """Probe process existence; fail closed toward 'alive' when unsure."""
+    try:
+        os.kill(int(pid), 0)
+        return True
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True
+    except Exception:
+        return True
+
+
+def assess_crashed_session_lease_orphan(
+    shadow: dict[str, Any] | None,
+    *,
+    lease: dict[str, Any] | None,
+    current_pid: int | None = None,
+    pid_alive: bool | None = None,
+) -> dict[str, Any]:
+    """Derive owner-process evidence for a durable lease from the shadow (#702).
+
+    Returns ``owner_process_alive`` tri-state. Only a provably dead recorded
+    owner pid yields ``False`` (a dead pid cannot still be running the owning
+    daemon); an alive pid is *not* ownership proof and stays ``None`` unless
+    it is this very process. Session ids must match exactly — the shadow of a
+    different lease proves nothing.
+    """
+    reasons: list[str] = []
+    if not shadow or not lease:
+        return {
+            "orphan_evidence": False,
+            "owner_process_alive": None,
+            "reasons": ["no durable session-lease shadow evidence available"],
+        }
+    shadow_sid = (shadow.get("session_id") or "").strip()
+    lease_sid = (lease.get("session_id") or "").strip()
+    if not shadow_sid or not lease_sid or shadow_sid != lease_sid:
+        return {
+            "orphan_evidence": False,
+            "owner_process_alive": None,
+            "reasons": [
+                "session-lease shadow session_id does not match the durable "
+                f"lease (shadow={shadow_sid or None}, lease={lease_sid or None})"
+            ],
+        }
+    owner_pid = shadow.get("owner_pid") or shadow.get("writer_pid")
+    try:
+        owner_pid = int(owner_pid)
+    except (TypeError, ValueError):
+        return {
+            "orphan_evidence": False,
+            "owner_process_alive": None,
+            "reasons": ["session-lease shadow has no usable owner pid (fail closed)"],
+        }
+    this_pid = current_pid if current_pid is not None else os.getpid()
+    if owner_pid == this_pid:
+        return {
+            "orphan_evidence": False,
+            "owner_process_alive": True,
+            "owner_pid": owner_pid,
+            "reasons": ["shadow owner pid is the current process"],
+        }
+    alive = pid_alive if pid_alive is not None else _pid_alive(owner_pid)
+    if alive:
+        reasons.append(
+            f"shadow owner pid {owner_pid} observed alive; PID liveness is not "
+            "ownership proof — owner evidence stays unknown (fail closed)"
+        )
+        return {
+            "orphan_evidence": False,
+            "owner_process_alive": None,
+            "owner_pid": owner_pid,
+            "reasons": reasons,
+        }
+    reasons.append(
+        f"shadow owner pid {owner_pid} is dead: the daemon that recorded the "
+        "matching session lease exited without teardown (crash orphan, #702)"
+    )
+    return {
+        "orphan_evidence": True,
+        "owner_process_alive": False,
+        "owner_pid": owner_pid,
+        "reasons": reasons,
+    }
+
+
 def assess_mutation_lease_gate(
     *,
     pr_number: int,
@@ -557,6 +733,7 @@ _HANDOFF_CLASSIFICATIONS = frozenset({
     "foreign_completed_superseded_head",
     "foreign_expired_superseded_head",
     "orphaned_owner_missing",
+    "orphaned_expired_superseded_head",
     "ambiguous_conflicting_evidence",
     "instructed_lease_missing_with_replacement",
     "worktree_binding_mismatch",
@@ -850,6 +1027,38 @@ def assess_obsolete_reviewer_comment_lease_cleanup(
             classification = "foreign_expired_superseded_head"
         elif head_superseded and has_terminal and not past_expiry:
             classification = "foreign_completed_superseded_head"
+        elif head_superseded and not has_terminal and past_expiry:
+            # Crash-orphan shape (#702): the lease outlived its head with no
+            # formal terminal review and has now passed canonical expiry, so
+            # it no longer gates fresh acquisition. Guarded cleanup of the
+            # ledger marker still demands provable owner-process-exit
+            # evidence and a safe worktree — never inferred.
+            classification = "orphaned_expired_superseded_head"
+            if owner_process_alive is True:
+                fail_closed_reasons.append(
+                    "lease past expiry on superseded head but owner process "
+                    "still alive; cleanup denied"
+                )
+            elif owner_process_alive is None:
+                fail_closed_reasons.append(
+                    "expired superseded-head lease with no terminal review: "
+                    "cleanup requires provable owner-process-exit evidence "
+                    "(owner_process_alive=false); the expired lease no longer "
+                    "gates fresh acquisition"
+                )
+            elif worktree_clean is False:
+                fail_closed_reasons.append(
+                    "owner process absent but worktree dirty; cleanup denied"
+                )
+            elif not (worktree_clean is True or worktree_exists is False):
+                # #702 F3: unknown worktree evidence must fail closed, exactly
+                # like the sibling orphaned_owner_missing gate and the
+                # diagnosis path — cleanup needs affirmative safe evidence.
+                fail_closed_reasons.append(
+                    "owner process absent but worktree evidence is unknown; "
+                    "cleanup requires affirmative safe evidence "
+                    "(worktree_clean=true or worktree_exists=false)"
+                )
         elif head_superseded and not has_terminal:
             classification = "ambiguous_conflicting_evidence"
             fail_closed_reasons.append(
@@ -891,6 +1100,7 @@ def assess_obsolete_reviewer_comment_lease_cleanup(
                 "foreign_expired_superseded_head",
                 "foreign_expired_current_head",
                 "orphaned_owner_missing",
+                "orphaned_expired_superseded_head",
             }:
                 fail_closed_reasons.append(
                     "controller/recovery capability required "
@@ -909,6 +1119,7 @@ def assess_obsolete_reviewer_comment_lease_cleanup(
             "foreign_completed_superseded_head",
             "foreign_expired_superseded_head",
             "orphaned_owner_missing",
+            "orphaned_expired_superseded_head",
             "foreign_expired_current_head",
         }
         # foreign_expired_current_head needs orphan/clean worktree + authority.
@@ -1181,6 +1392,30 @@ def diagnose_reviewer_pr_lease_handoff(
                 "foreign lease pinned to superseded head with completed formal "
                 "review; use guarded obsolete-lease cleanup (do not wait indefinitely)"
             )
+        elif not is_own and head_superseded and not terminal and past_expiry:
+            # Crash-orphan shape after canonical expiry (#702): the expired
+            # marker no longer gates acquisition, so a fresh reviewer may
+            # acquire at the current head. Guarded ledger cleanup becomes
+            # available only with provable owner-exit evidence.
+            classification = "orphaned_expired_superseded_head"
+            if owner_process_alive is False and (
+                worktree_clean is True or worktree_exists is False
+            ):
+                next_action = NEXT_ACTION_CLEANUP_OBSOLETE_LEASE
+                reasons.append(
+                    "expired superseded-head lease with no terminal review and "
+                    "provable owner-process exit (crash orphan, #702); guarded "
+                    "obsolete-lease cleanup eligible"
+                )
+            else:
+                next_action = NEXT_ACTION_ACQUIRE
+                reasons.append(
+                    "lease head superseded and lease past canonical expiry with "
+                    "no formal terminal review (crash orphan, #702); expired "
+                    "marker no longer gates acquisition — acquire a fresh lease "
+                    "at the current head; guarded cleanup needs "
+                    "owner_process_alive=false and a clean/absent worktree"
+                )
         elif not is_own and head_superseded and not terminal:
             classification = "ambiguous_conflicting_evidence"
             next_action = NEXT_ACTION_WAIT

stale_binding_recovery.py

@@ -0,0 +1,257 @@
+"""Sanctioned recovery for stale env-bound task workspace bindings (#702).
+
+When the MCP daemon dies from an unhandled exception it never runs teardown,
+and the auto-reconnected daemon inherits ``GITEA_ACTIVE_WORKTREE`` from its
+parent environment. That inherited value can permanently bind the fresh
+session to a prior task's worktree (the PR #701 / ``review-pr-654`` incident).
+
+This module classifies the active env binding against live session evidence
+and produces a fail-closed recovery plan. Only two situations permit the
+daemon to clear the binding on its own:
+
+- the bound path no longer exists on disk (``provably_stale_missing_path``)
+- the binding was inherited at daemon boot and a *sanctioned* in-session
+  reviewer/merger lease later bound a different worktree
+  (``superseded_by_session_lease``)
+
+Everything else is surfaced (``unverified_inherited``) and left for the
+managed reconnect path — never cleared silently, never rebound to a guessed
+worktree.
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Any
+
+ACTIVE_WORKTREE_ENV = "GITEA_ACTIVE_WORKTREE"
+
+CLASSIFICATION_UNBOUND = "unbound"
+CLASSIFICATION_CORROBORATED = "corroborated"
+CLASSIFICATION_UNVERIFIED_INHERITED = "unverified_inherited"
+CLASSIFICATION_PROVABLY_STALE_MISSING_PATH = "provably_stale_missing_path"
+CLASSIFICATION_SUPERSEDED_BY_SESSION_LEASE = "superseded_by_session_lease"
+
+CLASSIFICATIONS = frozenset({
+    CLASSIFICATION_UNBOUND,
+    CLASSIFICATION_CORROBORATED,
+    CLASSIFICATION_UNVERIFIED_INHERITED,
+    CLASSIFICATION_PROVABLY_STALE_MISSING_PATH,
+    CLASSIFICATION_SUPERSEDED_BY_SESSION_LEASE,
+})
+
+# Roles whose task workspace is owned by a lease lifecycle, so an inherited
+# generic binding without a corroborating lease is suspect.
+LEASE_BOUND_ROLES = frozenset({"reviewer", "merger"})
+
+RECOVERY_ACTION_NONE = "none"
+RECOVERY_ACTION_CLEAR_ENV = "clear_env"
+
+
+def _norm(path: str | None) -> str:
+    text = (path or "").strip()
+    if not text:
+        return ""
+    return os.path.realpath(os.path.abspath(text))
+
+
+def snapshot_boot_bindings(
+    env: dict[str, str] | os._Environ | None = None,
+) -> dict[str, Any]:
+    """Capture worktree env bindings present when the daemon booted.
+
+    A value present in this snapshot was inherited from the parent
+    environment, not established by any sanctioned tool in this daemon's
+    lifetime.
+    """
+    env_map = env if env is not None else os.environ
+    return {
+        "active_worktree": (env_map.get(ACTIVE_WORKTREE_ENV) or "").strip() or None,
+    }
+
+
+def classify_active_worktree_binding(
+    *,
+    active_value: str | None,
+    role_env_value: str | None = None,
+    session_lease_worktree: str | None = None,
+    boot_inherited: bool,
+    path_exists: bool | None,
+    role_kind: str | None = None,
+) -> dict[str, Any]:
+    """Classify the GITEA_ACTIVE_WORKTREE binding against live evidence.
+
+    Fail closed: unknown evidence never yields a clear-eligible class.
+    """
+    reasons: list[str] = []
+    active = _norm(active_value)
+    if not active:
+        return {
+            "classification": CLASSIFICATION_UNBOUND,
+            "clear_eligible": False,
+            "active_worktree": None,
+            "reasons": ["no GITEA_ACTIVE_WORKTREE binding present"],
+        }
+
+    role = (role_kind or "").strip().lower()
+    role_env = _norm(role_env_value)
+    lease_wt = _norm(session_lease_worktree)
+
+    result: dict[str, Any] = {
+        "active_worktree": active,
+        "boot_inherited": bool(boot_inherited),
+        "role_kind": role or None,
+        "session_lease_worktree": lease_wt or None,
+        "role_env_worktree": role_env or None,
+    }
+
+    if path_exists is False:
+        reasons.append(
+            f"bound worktree '{active}' no longer exists on disk; the binding "
+            "cannot name a valid task workspace"
+        )
+        result.update({
+            "classification": CLASSIFICATION_PROVABLY_STALE_MISSING_PATH,
+            "clear_eligible": True,
+            "reasons": reasons,
+        })
+        return result
+
+    if lease_wt and lease_wt == active:
+        reasons.append("binding matches the sanctioned in-session lease worktree")
+        result.update({
+            "classification": CLASSIFICATION_CORROBORATED,
+            "clear_eligible": False,
+            "reasons": reasons,
+        })
+        return result
+
+    if lease_wt and lease_wt != active and boot_inherited:
+        reasons.append(
+            "boot-inherited binding disagrees with the sanctioned in-session "
+            f"lease worktree '{lease_wt}'; the lease is live authority"
+        )
+        result.update({
+            "classification": CLASSIFICATION_SUPERSEDED_BY_SESSION_LEASE,
+            "clear_eligible": True,
+            "reasons": reasons,
+        })
+        return result
+
+    if lease_wt and lease_wt != active and not boot_inherited:
+        # Set after boot by tooling; disagreement is surfaced, never auto-cleared.
+        reasons.append(
+            "post-boot binding disagrees with the in-session lease worktree; "
+            "surfacing only (fail closed, no automatic clear)"
+        )
+        result.update({
+            "classification": CLASSIFICATION_UNVERIFIED_INHERITED,
+            "clear_eligible": False,
+            "reasons": reasons,
+        })
+        return result
+
+    if role_env and role_env == active:
+        reasons.append("binding matches the role-specific worktree env binding")
+        result.update({
+            "classification": CLASSIFICATION_CORROBORATED,
+            "clear_eligible": False,
+            "reasons": reasons,
+        })
+        return result
+
+    if boot_inherited and role in LEASE_BOUND_ROLES and not lease_wt:
+        reasons.append(
+            "binding was inherited at daemon boot, no sanctioned in-session "
+            f"lease corroborates it, and the {role} role binds workspaces "
+            "through the lease lifecycle; treat as unverified until a fresh "
+            "lease or managed reconnect re-establishes the workspace"
+        )
+        result.update({
+            "classification": CLASSIFICATION_UNVERIFIED_INHERITED,
+            "clear_eligible": False,
+            "reasons": reasons,
+        })
+        return result
+
+    reasons.append("no contradicting evidence; binding treated as corroborated")
+    result.update({
+        "classification": CLASSIFICATION_CORROBORATED,
+        "clear_eligible": False,
+        "reasons": reasons,
+    })
+    return result
+
+
+def plan_recovery(classification_result: dict[str, Any]) -> dict[str, Any]:
+    """Turn a classification into an explicit, fail-closed recovery plan."""
+    classification = classification_result.get("classification")
+    clear_eligible = bool(classification_result.get("clear_eligible"))
+    reasons = list(classification_result.get("reasons") or [])
+
+    if classification not in CLASSIFICATIONS:
+        return {
+            "action": RECOVERY_ACTION_NONE,
+            "clear_allowed": False,
+            "classification": classification,
+            "reasons": reasons + ["unknown classification (fail closed)"],
+        }
+
+    if clear_eligible and classification in {
+        CLASSIFICATION_PROVABLY_STALE_MISSING_PATH,
+        CLASSIFICATION_SUPERSEDED_BY_SESSION_LEASE,
+    }:
+        return {
+            "action": RECOVERY_ACTION_CLEAR_ENV,
+            "clear_allowed": True,
+            "classification": classification,
+            "active_worktree": classification_result.get("active_worktree"),
+            "reasons": reasons,
+        }
+
+    exact_next_action = None
+    if classification == CLASSIFICATION_UNVERIFIED_INHERITED:
+        exact_next_action = (
+            "managed recovery required: reconnect the MCP namespace through "
+            "the managed client configuration (or start a fresh session) so "
+            "the daemon boots without the inherited GITEA_ACTIVE_WORKTREE, or "
+            "corroborate the binding by acquiring a lease for that worktree; "
+            "do not clear or rewrite the environment manually"
+        )
+    return {
+        "action": RECOVERY_ACTION_NONE,
+        "clear_allowed": False,
+        "classification": classification,
+        "active_worktree": classification_result.get("active_worktree"),
+        "reasons": reasons,
+        **({"exact_next_action": exact_next_action} if exact_next_action else {}),
+    }
+
+
+def apply_recovery(
+    plan: dict[str, Any],
+    env: dict[str, str] | os._Environ | None = None,
+) -> dict[str, Any]:
+    """Apply a clear-eligible plan by removing the stale env binding.
+
+    This is the sanctioned in-daemon mechanism (#702 AC2); callers must
+    persist the returned record for audit. Refuses anything but an explicit
+    ``clear_env`` plan.
+    """
+    env_map = env if env is not None else os.environ
+    if not plan.get("clear_allowed") or plan.get("action") != RECOVERY_ACTION_CLEAR_ENV:
+        return {
+            "performed": False,
+            "action": RECOVERY_ACTION_NONE,
+            "reasons": ["recovery plan does not authorize clearing (fail closed)"],
+        }
+    cleared_value = (env_map.get(ACTIVE_WORKTREE_ENV) or "").strip() or None
+    env_map.pop(ACTIVE_WORKTREE_ENV, None)
+    return {
+        "performed": True,
+        "action": RECOVERY_ACTION_CLEAR_ENV,
+        "cleared_env": ACTIVE_WORKTREE_ENV,
+        "cleared_value": cleared_value,
+        "classification": plan.get("classification"),
+        "reasons": list(plan.get("reasons") or []),
+    }

stale_review_decision_lock.py

@@ -438,3 +438,252 @@ def format_cleanup_audit_comment(audit: dict[str, Any]) -> str:
         "This path only clears a lock when the referenced PR is merged/closed.",
     ]
     return "\n".join(lines)
+
+
+# ---------------------------------------------------------------------------
+# #709 cross-profile cleanup, overwrite protection, recovery provenance
+# ---------------------------------------------------------------------------
+
+
+def has_unresolved_terminal_evidence(lock: dict | None) -> bool:
+    """True when *lock* still carries a terminal live mutation ledger."""
+    return last_terminal_mutation(lock) is not None
+
+
+def assess_init_overwrite(
+    existing_lock: dict | None,
+    *,
+    force: bool = False,
+) -> dict[str, Any]:
+    """Whether init_review_decision_lock may replace an existing durable lock (#709 AC2).
+
+    Unresolved terminal evidence must never be silently replaced by an empty
+    initialized lock. *force* does not authorize destruction of terminal
+    ledgers — only explicit cleanup / archive transitions may remove them.
+    """
+    result: dict[str, Any] = {
+        "overwrite_allowed": True,
+        "has_existing": existing_lock is not None,
+        "has_unresolved_terminal": False,
+        "reasons": [],
+        "last_terminal_pr": None,
+        "last_terminal_action": None,
+        "existing_profile_identity": None,
+    }
+    if existing_lock is None:
+        result["reasons"].append("no existing lock; empty init allowed")
+        return result
+    result["existing_profile_identity"] = (
+        existing_lock.get("profile_identity")
+        or existing_lock.get("session_profile_lock")
+        or existing_lock.get("session_profile")
+    )
+    last = last_terminal_mutation(existing_lock)
+    if last is None:
+        result["reasons"].append(
+            "existing lock has no terminal mutations; re-init allowed"
+        )
+        return result
+    result["has_unresolved_terminal"] = True
+    result["overwrite_allowed"] = False
+    result["last_terminal_pr"] = last.get("pr_number")
+    result["last_terminal_action"] = last.get("action")
+    result["reasons"].append(
+        "refuse empty re-init: unresolved terminal decision-lock evidence "
+        f"present ({last.get('action')} on PR #{last.get('pr_number')}); "
+        "use gitea_cleanup_stale_review_decision_lock when moot, or archive "
+        f"via sanctioned recovery — force={force!r} does not authorize overwrite "
+        "(#709 AC2)"
+    )
+    return result
+
+
+def lock_targets_merged_pr_approval(
+    lock: dict | None,
+    *,
+    pr_number: int,
+    expected_head_sha: str | None = None,
+) -> bool:
+    """True when *lock*'s last terminal mutation is approve of *pr_number*.
+
+    When *expected_head_sha* is provided, a **recorded** terminal head must
+    match. Legacy same-repo approve locks with no recorded head do **not**
+    match (fail closed, #709 F3 residual / review 435) — callers must use the
+    strict secondary path that refuses no-head clears.
+    """
+    last = last_terminal_mutation(lock)
+    if last is None:
+        return False
+    if last.get("action") != "approve":
+        return False
+    if last.get("pr_number") != pr_number:
+        return False
+    if expected_head_sha:
+        want = normalize_head_sha(expected_head_sha)
+        if not want:
+            return False
+        locked = mutation_head_sha(last, lock)
+        # Require recorded-head match; unrecorded head is not a match (#709 F3).
+        if not locked or not heads_equal(locked, want):
+            return False
+    return True
+
+
+def build_post_merge_recovery_record(
+    *,
+    pr_number: int,
+    head_sha: str | None,
+    merge_commit_sha: str | None,
+    target_profile_identity: str | None,
+    failed_step: str,
+    error: str | None,
+    remote: str | None,
+    org: str | None,
+    repo: str | None,
+    actor_username: str | None,
+    profile_name: str | None,
+) -> dict[str, Any]:
+    """Durable recovery-required payload after irreversible merge (#709 AC3)."""
+    return {
+        "event": "post_merge_decision_lock_recovery_required",
+        "status": "recovery_required",
+        "issue_ref": "#709",
+        "recovery_critical": True,
+        "applied": False,  # never claim historical cleanup succeeded
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "pr_number": pr_number,
+        "head_sha": normalize_head_sha(head_sha),
+        "merge_commit_sha": merge_commit_sha,
+        "target_profile_identity": target_profile_identity,
+        "failed_step": failed_step,
+        "error": error,
+        "remote": remote,
+        "org": org,
+        "repo": repo,
+        "actor_username": actor_username,
+        "profile_name": profile_name,
+        "required_recovery_action": (
+            "retry cross-profile decision-lock cleanup and audit publication "
+            "for the merged PR; if terminal evidence is gone, use "
+            "gitea_record_irrecoverable_decision_lock_provenance"
+        ),
+    }
+
+
+def build_irrecoverable_provenance_record(
+    *,
+    pr_number: int,
+    head_sha: str | None,
+    remote: str | None,
+    org: str | None,
+    repo: str | None,
+    actor_username: str | None,
+    profile_name: str | None,
+    reason: str,
+    incident_ref: str | None = None,
+    # Deprecated kwargs retained only so stale call sites fail closed:
+    operator_authorized: bool | None = None,
+    # Required for merger-acceptable records (#709 F1):
+    authorization: dict[str, Any] | None = None,
+    incident_issue: int | None = None,
+    incident_comment_id: int | None = None,
+    destroyed_subject: str | None = None,
+    historical_provenance_subject: str | None = None,
+) -> dict[str, Any]:
+    """Truthful absence-of-proof record (#709 AC5). Never sets applied=True.
+
+    Caller-supplied ``operator_authorized`` is **ignored** as authorization
+    evidence (review 434 F1). Prefer
+    :func:`irrecoverable_provenance.build_irrecoverable_provenance_record`
+    with a verified server-side authorization artifact.
+    """
+    # Explicitly ignore deprecated self-assertable Boolean.
+    _ = operator_authorized
+    if authorization is not None and incident_issue is not None and incident_comment_id is not None:
+        from irrecoverable_provenance import (
+            build_irrecoverable_provenance_record as _build,
+        )
+
+        return _build(
+            pr_number=int(pr_number),
+            head_sha=str(head_sha or ""),
+            remote=str(remote or ""),
+            org=str(org or ""),
+            repo=str(repo or ""),
+            actor_username=actor_username,
+            profile_name=profile_name,
+            reason=reason,
+            incident_issue=int(incident_issue),
+            incident_comment_id=int(incident_comment_id),
+            authorization=authorization,
+            destroyed_subject=destroyed_subject,
+            historical_provenance_subject=historical_provenance_subject
+            or destroyed_subject,
+        )
+    # Fail-closed skeleton when no server authorization is supplied: never
+    # sets merger_may_accept True (even if operator_authorized was True).
+    return {
+        "event": "irrecoverable_decision_lock_provenance",
+        "status": "provenance_irrecoverable",
+        "record_type": "irrecoverable_decision_provenance",
+        "operator_recovery_required": True,
+        "issue_ref": "#709",
+        "recovery_critical": True,
+        "applied": False,
+        "historical_cleanup_proven": False,
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "pr_number": pr_number,
+        "head_sha": normalize_head_sha(head_sha),
+        "remote": remote,
+        "org": org,
+        "repo": repo,
+        "actor_username": actor_username,
+        "profile_name": profile_name,
+        "reason": reason,
+        "incident_ref": incident_ref,
+        "incident_issue": incident_issue,
+        "incident_comment_id": incident_comment_id,
+        "authorization_verified": False,
+        "merger_may_accept": False,
+        "acceptance_rule": (
+            "Merger may accept this record only when a server-side "
+            "authorization artifact verifies for remote/org/repo/PR/exact "
+            "head/incident, the record is durable and read back, and normal "
+            "merge gates still pass. Caller Booleans never authorize. This "
+            "does not prove historical cleanup."
+        ),
+    }
+
+
+def format_irrecoverable_audit_comment(record: dict[str, Any]) -> str:
+    """Markdown body for irrecoverable provenance audit (no applied=true claim)."""
+    from irrecoverable_provenance import (
+        format_irrecoverable_audit_comment as _fmt,
+    )
+
+    return _fmt(record)
+
+
+def format_post_merge_recovery_comment(record: dict[str, Any]) -> str:
+    """Markdown body for post-merge recovery-required audit."""
+    lines = [
+        "## Post-merge decision-lock recovery required (#709)",
+        "",
+        "Status: **RECOVERY_REQUIRED** (merge is irreversible; cleanup/audit incomplete)",
+        "",
+        f"- actor: `{record.get('actor_username')}`",
+        f"- profile: `{record.get('profile_name')}`",
+        f"- timestamp: `{record.get('timestamp')}`",
+        f"- PR: `#{record.get('pr_number')}`",
+        f"- head_sha: `{record.get('head_sha')}`",
+        f"- merge_commit_sha: `{record.get('merge_commit_sha')}`",
+        f"- target_profile_identity: `{record.get('target_profile_identity')}`",
+        f"- failed_step: `{record.get('failed_step')}`",
+        f"- error: `{record.get('error')}`",
+        "",
+        f"Required action: {record.get('required_recovery_action')}",
+        "",
+        "applied=false — do not treat this as successful cleanup evidence.",
+    ]
+    return "\n".join(lines)
+

task_capability_map.py

@@ -84,7 +84,7 @@ TASK_CAPABILITY_MAP: dict[str, dict[str, str]] = {
     },
     "adopt_merger_pr_lease": {
         "permission": "gitea.pr.comment",
-        "role": "reviewer",
+        "role": "merger",
     },
     # #691: guarded non-owner cleanup of obsolete comment-backed reviewer leases.
     # Apply path posts lease release + audit comments (gitea.pr.comment).
@@ -127,6 +127,32 @@ TASK_CAPABILITY_MAP: dict[str, dict[str, str]] = {
         "permission": "gitea.pr.review",
         "role": "reviewer",
     },
+    # #709: truthful absence-of-proof recovery (server-side auth + record + consume).
+    # Dedicated mutation capability — gitea.read is insufficient (review 434 F1).
+    "issue_irrecoverable_provenance_authorization": {
+        "permission": "gitea.decision_lock.irrecoverable_recovery",
+        "role": "reconciler",
+    },
+    "gitea_issue_irrecoverable_provenance_authorization": {
+        "permission": "gitea.decision_lock.irrecoverable_recovery",
+        "role": "reconciler",
+    },
+    "record_irrecoverable_decision_lock_provenance": {
+        "permission": "gitea.decision_lock.irrecoverable_recovery",
+        "role": "reconciler",
+    },
+    "gitea_record_irrecoverable_decision_lock_provenance": {
+        "permission": "gitea.decision_lock.irrecoverable_recovery",
+        "role": "reconciler",
+    },
+    "consume_irrecoverable_decision_lock_provenance": {
+        "permission": "gitea.pr.merge",
+        "role": "merger",
+    },
+    "gitea_consume_irrecoverable_decision_lock_provenance": {
+        "permission": "gitea.pr.merge",
+        "role": "merger",
+    },
     "delete_branch": {
         "permission": "gitea.branch.delete",
         "role": "author",

tests/test_api_reliability.py

@@ -79,41 +79,46 @@ class TestApiRequestFailures(unittest.TestCase):
     @patch("gitea_auth.urllib.request.urlopen")
     def test_timeout_converted_to_runtimeerror(self, mock_open):
         mock_open.side_effect = TimeoutError("timed out")
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaNetworkError) as ctx:
             gitea_auth.api_request("GET", URL, FAKE_AUTH)
-        self.assertIn("network error contacting Gitea", str(ctx.exception))
+        # Fixed message only (#699) — still a RuntimeError subclass.
+        self.assertIsInstance(ctx.exception, RuntimeError)
+        self.assertEqual(str(ctx.exception), "Network error contacting Gitea")
 
     @patch("gitea_auth.urllib.request.urlopen")
     def test_dns_network_failure_converted(self, mock_open):
         mock_open.side_effect = urllib.error.URLError("Name or service not known")
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaNetworkError) as ctx:
             gitea_auth.api_request("GET", URL, FAKE_AUTH)
-        self.assertIn("network error contacting Gitea", str(ctx.exception))
+        self.assertEqual(str(ctx.exception), "Network error contacting Gitea")
 
     @patch("gitea_auth.urllib.request.urlopen")
     def test_502_upstream_message(self, mock_open):
         mock_open.side_effect = http_error(502, "bad gateway")
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
             gitea_auth.api_request("GET", URL, FAKE_AUTH)
         msg = str(ctx.exception)
-        self.assertIn("HTTP 502", msg)
+        self.assertEqual(msg, "Gitea upstream unavailable")
-        self.assertIn("upstream unavailable", msg)
+        self.assertEqual(ctx.exception.reason_code, "upstream_unavailable")
+        self.assertEqual(ctx.exception.http_status, 502)
 
     @patch("gitea_auth.urllib.request.urlopen")
     def test_503_upstream_message(self, mock_open):
         mock_open.side_effect = http_error(503, "")
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
             gitea_auth.api_request("GET", URL, FAKE_AUTH)
-        self.assertIn("HTTP 503", str(ctx.exception))
+        self.assertEqual(str(ctx.exception), "Gitea upstream unavailable")
-        self.assertIn("upstream unavailable", str(ctx.exception))
+        self.assertEqual(ctx.exception.http_status, 503)
 
     @patch("gitea_auth.urllib.request.urlopen")
     def test_malformed_error_payload_does_not_crash(self, mock_open):
-        # Non-JSON garbage error body must still yield a clean RuntimeError.
+        # Non-JSON garbage error body must still yield a clean typed error
+        # with a fixed message (no body echo — #699).
         mock_open.side_effect = http_error(500, "<html>garbage</html>")
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
             gitea_auth.api_request("GET", URL, FAKE_AUTH)
-        self.assertIn("HTTP 500", str(ctx.exception))
+        self.assertEqual(str(ctx.exception), "Gitea HTTP request failed")
+        self.assertNotIn("<html>", str(ctx.exception))
 
     @patch("gitea_auth.urllib.request.urlopen")
     def test_malformed_success_json_raises_clean_error(self, mock_open):
@@ -126,16 +131,17 @@ class TestApiRequestFailures(unittest.TestCase):
     def test_no_secret_leak_in_error_body(self, mock_open):
         mock_open.side_effect = http_error(
             400, "failed: token supersecret123 rejected")
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
             gitea_auth.api_request("GET", URL, FAKE_AUTH)
         msg = str(ctx.exception)
         self.assertNotIn("supersecret123", msg)
-        self.assertIn(gitea_audit.REDACTED, msg)
+        # Fixed message — body never appears (stronger than redaction).
+        self.assertEqual(msg, "Gitea HTTP request failed")
 
     @patch("gitea_auth.urllib.request.urlopen")
     def test_auth_header_never_in_error(self, mock_open):
         mock_open.side_effect = http_error(400, "bad request")
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
             gitea_auth.api_request("GET", URL, FAKE_AUTH)
         self.assertNotIn(FAKE_AUTH, str(ctx.exception))
 

tests/test_audit_reconciliation_mode.py

@@ -27,7 +27,13 @@ from task_capability_map import required_permission, required_role
 
 DELETE_PROFILE = {
     "profile_name": "prgs-author-delete",
-    "allowed_operations": ["gitea.read", "gitea.branch.delete"],
+    "role": "author",
+    "allowed_operations": [
+        "gitea.read",
+        "gitea.pr.create",
+        "gitea.branch.push",
+        "gitea.branch.delete",
+    ],
     "forbidden_operations": [],
     "audit_label": "prgs-author-delete",
 }

tests/test_issue_691_obsolete_reviewer_lease_cleanup.py

@@ -561,12 +561,29 @@ class TestIssue691SupersededHeadCleanup(unittest.TestCase):
         self.assertTrue(any("owns the lease" in r for r in result["reasons"]))
 
     def test_superseded_without_terminal_review_denied(self):
+        # Pre-expiry the missing terminal review is ambiguous (fail closed);
+        # post-expiry with unknown owner evidence it is a crash-orphan (#702)
+        # that still fails closed until owner-process exit is proven.
         now = datetime(2026, 7, 13, 5, 0, 0, tzinfo=timezone.utc)
         expires = datetime(2026, 7, 13, 4, 23, 0, tzinfo=timezone.utc)
         claim = _lease_comment(expires_at=expires, last_activity=expires - timedelta(hours=1))
-        result = _assess([claim], formal_reviews=[], now=now)
+        result = _assess(
+            [claim], formal_reviews=[], now=now, owner_process_alive=None
+        )
         self.assertFalse(result["cleanup_allowed"])
-        self.assertEqual(result["classification"], "ambiguous_conflicting_evidence")
+        self.assertEqual(
+            result["classification"], "orphaned_expired_superseded_head"
+        )
+
+        fresh_expires = now + timedelta(minutes=60)
+        fresh = _lease_comment(
+            expires_at=fresh_expires, last_activity=now - timedelta(minutes=5)
+        )
+        pre_expiry = _assess([fresh], formal_reviews=[], now=now)
+        self.assertFalse(pre_expiry["cleanup_allowed"])
+        self.assertEqual(
+            pre_expiry["classification"], "ambiguous_conflicting_evidence"
+        )
 
 
 class TestIssue691DiagnoseClassifications(unittest.TestCase):

tests/test_issue_702_review_findings_f1_f6.py

@@ -0,0 +1,693 @@
+"""Regression tests for the PR #703 REQUEST_CHANGES findings F1–F6 (#702).
+
+Formal review at head 889931d independently reproduced six defects:
+
+F1 — stale-binding recovery ran only AFTER the terminal launcher probe in
+     ``gitea_resolve_task_capability``; a worktree removed mid-session wedged
+     every mutation-task resolution on 'missing cwd' before recovery.
+F2 — ``_resolve_preflight_workspace_path`` skipped the existence checks the
+     canonical mutation context applies; a dead binding could produce empty
+     porcelain and read as a clean workspace.
+F3 — ``orphaned_expired_superseded_head`` cleanup accepted unknown
+     ``worktree_exists``/``worktree_clean`` evidence.
+F4 — the 4-hour session-state TTL silently discarded recovery-critical
+     crash-orphan shadow evidence.
+F5 — the single same-profile shadow slot let concurrent sessions overwrite
+     each other's crash evidence.
+F6 — the environment was cleared BEFORE the audit record was persisted, and
+     audit-write failure was swallowed.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import shutil
+import sys
+import tempfile
+import unittest
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+from unittest.mock import patch
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+import mcp_session_state as mss  # noqa: E402
+import namespace_workspace_binding as nwb  # noqa: E402
+import reviewer_pr_lease as leases  # noqa: E402
+import stale_binding_recovery as sbr  # noqa: E402
+import mcp_server  # noqa: E402
+
+REPO = "Scaled-Tech-Consulting/Gitea-Tools"
+OLD_HEAD = "b" * 40
+NEW_HEAD = "6" * 40
+
+CONFIG = {
+    "version": 2,
+    "contexts": {
+        "ctx": {
+            "enabled": True,
+            "gitea": {"enabled": True, "base_url": "https://gitea.example.com"},
+        }
+    },
+    "profiles": {
+        "prgs-author": {
+            "enabled": True,
+            "context": "ctx",
+            "role": "author",
+            "username": "jcwalker3",
+            "auth": {"type": "env", "name": "GITEA_TOKEN_AUTHOR"},
+            "allowed_operations": [
+                "gitea.read", "gitea.issue.create", "gitea.issue.comment",
+                "gitea.pr.create", "gitea.branch.push",
+            ],
+            "forbidden_operations": [
+                "gitea.pr.approve", "gitea.pr.merge", "gitea.pr.review",
+            ],
+            "execution_profile": "prgs-author",
+        },
+    },
+    "rules": {"allow_runtime_switching": False},
+}
+
+
+def _now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+class _CapabilityHarness(unittest.TestCase):
+    """Shared config/env plumbing for resolve-capability integration tests."""
+
+    def setUp(self):
+        self._remotes = patch.dict(mcp_server.REMOTES, {
+            "prgs": {
+                "host": "gitea.example.com",
+                "org": "Scaled-Tech-Consulting",
+                "repo": "Gitea-Tools",
+            },
+        })
+        self._remotes.start()
+        mcp_server._IDENTITY_CACHE.clear()
+        self._dir = tempfile.TemporaryDirectory()
+        self.config_path = os.path.join(self._dir.name, "profiles.json")
+        with open(self.config_path, "w", encoding="utf-8") as fh:
+            fh.write(json.dumps(CONFIG))
+
+    def tearDown(self):
+        self._remotes.stop()
+        mcp_server._IDENTITY_CACHE.clear()
+        self._dir.cleanup()
+
+    def _env(self, **extra):
+        env = {
+            "GITEA_MCP_CONFIG": self.config_path,
+            "GITEA_MCP_PROFILE": "prgs-author",
+            "GITEA_TOKEN_AUTHOR": "author-pass",
+        }
+        env.update(extra)
+        return env
+
+    def _deleted_worktree(self) -> str:
+        path = tempfile.mkdtemp(prefix="gitea-removed-worktree-")
+        shutil.rmtree(path)
+        return path
+
+
+class TestF1RecoveryBeforeTerminalProbe(_CapabilityHarness):
+    """F1: recovery must run before the terminal cwd probe can wedge."""
+
+    @patch("mcp_server.api_request", return_value={"login": "jcwalker3"})
+    @patch("mcp_server.get_auth_header", return_value="token author-pass")
+    def test_removed_worktree_recovers_before_probe(self, _auth, _api):
+        missing = self._deleted_worktree()
+        env = self._env(
+            GITEA_ACTIVE_WORKTREE=missing,
+            GITEA_FORCE_TERMINAL_PROBE="1",
+            GITEA_FORCE_STALE_BINDING_RECOVERY="1",
+        )
+        with patch.dict(os.environ, env):
+            res = mcp_server.gitea_resolve_task_capability(
+                task="create_pr", remote="prgs"
+            )
+            # The provably-stale binding was cleared before the probe ran.
+            self.assertNotIn("GITEA_ACTIVE_WORKTREE", os.environ)
+        self.assertFalse(res.get("terminal_launcher_unhealthy"))
+        report = res.get("stale_binding_recovery") or {}
+        self.assertEqual(
+            report.get("classification"),
+            sbr.CLASSIFICATION_PROVABLY_STALE_MISSING_PATH,
+        )
+        self.assertTrue(report.get("recovery_performed"))
+        self.assertTrue(res.get("allowed_in_current_session"))
+
+    @patch("mcp_server.api_request", return_value={"login": "jcwalker3"})
+    @patch("mcp_server.get_auth_header", return_value="token author-pass")
+    def test_probe_block_still_carries_recovery_report(self, _auth, _api):
+        # Recovery disabled (test mode without the force flag) preserves the
+        # fail-closed probe block, but the block must now carry the
+        # classification evidence instead of hiding it.
+        missing = self._deleted_worktree()
+        env = self._env(
+            GITEA_ACTIVE_WORKTREE=missing,
+            GITEA_FORCE_TERMINAL_PROBE="1",
+        )
+        with patch.dict(os.environ, env):
+            res = mcp_server.gitea_resolve_task_capability(
+                task="create_pr", remote="prgs"
+            )
+            self.assertIn("GITEA_ACTIVE_WORKTREE", os.environ)
+        self.assertTrue(res.get("terminal_launcher_unhealthy"))
+        report = res.get("stale_binding_recovery") or {}
+        self.assertEqual(
+            report.get("classification"),
+            sbr.CLASSIFICATION_PROVABLY_STALE_MISSING_PATH,
+        )
+        self.assertFalse(report.get("recovery_performed"))
+
+    @patch("mcp_server.api_request", return_value={"login": "jcwalker3"})
+    @patch("mcp_server.get_auth_header", return_value="token author-pass")
+    def test_unverified_binding_is_never_cleared_by_reordering(self, _auth, _api):
+        # F1 must not weaken authorization: an existing, uncorroborated
+        # binding survives resolution untouched even with recovery forced.
+        existing = tempfile.mkdtemp(prefix="gitea-existing-worktree-")
+        self.addCleanup(shutil.rmtree, existing, ignore_errors=True)
+        env = self._env(
+            GITEA_ACTIVE_WORKTREE=existing,
+            GITEA_FORCE_TERMINAL_PROBE="1",
+            GITEA_FORCE_STALE_BINDING_RECOVERY="1",
+        )
+        with patch.dict(os.environ, env):
+            res = mcp_server.gitea_resolve_task_capability(
+                task="create_pr", remote="prgs"
+            )
+            self.assertEqual(os.environ.get("GITEA_ACTIVE_WORKTREE"), existing)
+        report = res.get("stale_binding_recovery") or {}
+        self.assertFalse(report.get("recovery_performed"))
+
+
+class TestF2PreflightPathVerification(_CapabilityHarness):
+    """F2: preflight and mutation contexts verify resolved paths alike."""
+
+    def test_preflight_and_mutation_context_agree_on_dead_binding(self):
+        missing = self._deleted_worktree()
+        env = self._env(GITEA_ACTIVE_WORKTREE=missing)
+        with patch.dict(os.environ, env):
+            preflight = mcp_server._resolve_preflight_workspace_path(None)
+            mutation = mcp_server._resolve_namespace_mutation_context(None)
+        self.assertEqual(preflight, mutation["workspace_path"])
+        self.assertNotEqual(preflight, os.path.realpath(missing))
+
+    def test_missing_explicit_worktree_never_reads_clean(self):
+        missing = self._deleted_worktree()
+        with patch.dict(os.environ, self._env()):
+            porcelain = mcp_server._get_workspace_porcelain(worktree_path=missing)
+        self.assertEqual(
+            porcelain, mcp_server.PREFLIGHT_WORKSPACE_UNAVAILABLE_PORCELAIN
+        )
+        # The sentinel parses as a tracked dirty entry — never clean.
+        self.assertTrue(mcp_server._parse_porcelain_entries(porcelain))
+
+    def test_workspace_deleted_during_evaluation_never_reads_clean(self):
+        # TOCTOU: resolution succeeded, then the path vanished before git ran.
+        missing = self._deleted_worktree()
+        with patch.dict(os.environ, self._env()):
+            with patch.object(
+                mcp_server,
+                "_resolve_preflight_workspace_path",
+                return_value=missing,
+            ):
+                porcelain = mcp_server._get_workspace_porcelain()
+        self.assertEqual(
+            porcelain, mcp_server.PREFLIGHT_WORKSPACE_UNAVAILABLE_PORCELAIN
+        )
+
+    def test_git_failure_never_reads_clean(self):
+        class _Failed:
+            returncode = 128
+            stdout = ""
+            stderr = "fatal: not a git repository"
+
+        with patch.dict(os.environ, self._env()):
+            with patch.object(
+                mcp_server.subprocess, "run", return_value=_Failed()
+            ):
+                porcelain = mcp_server._get_workspace_porcelain()
+        self.assertEqual(
+            porcelain, mcp_server.PREFLIGHT_WORKSPACE_UNAVAILABLE_PORCELAIN
+        )
+
+    def test_symlinked_binding_to_deleted_target_is_demoted(self):
+        target = tempfile.mkdtemp(prefix="gitea-symlink-target-")
+        holder = tempfile.mkdtemp(prefix="gitea-symlink-holder-")
+        self.addCleanup(shutil.rmtree, holder, ignore_errors=True)
+        link = os.path.join(holder, "worktree-link")
+        os.symlink(target, link)
+        shutil.rmtree(target)
+        demotions: list[str] = []
+        _path, source = nwb.resolve_namespace_workspace(
+            role_kind="reviewer",
+            process_project_root=os.getcwd(),
+            env={"GITEA_ACTIVE_WORKTREE": link},
+            demotions=demotions,
+            verify_paths=True,
+        )
+        self.assertEqual(source, "MCP server process root (default)")
+        self.assertEqual(len(demotions), 1)
+
+    def test_dead_binding_falls_through_to_live_role_binding(self):
+        # Path changes during evaluation: the dead candidate demotes at
+        # verification time and never resurrects over the live fallback.
+        alive = tempfile.mkdtemp(prefix="gitea-alive-worktree-")
+        self.addCleanup(shutil.rmtree, alive, ignore_errors=True)
+        missing = self._deleted_worktree()
+        path, source = nwb.resolve_namespace_workspace(
+            role_kind="reviewer",
+            process_project_root=os.getcwd(),
+            env={
+                "GITEA_ACTIVE_WORKTREE": missing,
+                "GITEA_REVIEWER_WORKTREE": alive,
+            },
+            verify_paths=True,
+        )
+        self.assertEqual(path, os.path.realpath(alive))
+        self.assertEqual(
+            source, "GITEA_REVIEWER_WORKTREE environment variable"
+        )
+
+
+class TestF3AffirmativeWorktreeEvidence(unittest.TestCase):
+    """F3: unknown worktree evidence must fail closed for crash orphans."""
+
+    def _comments(self) -> list[dict]:
+        stale = _now() - timedelta(hours=3)
+        body = leases.format_lease_body(
+            repo=REPO,
+            pr_number=701,
+            issue_number=699,
+            reviewer_identity="sysadmin",
+            profile="prgs-reviewer",
+            session_id="65664-4f248d213d60",
+            worktree="branches/review-pr-654",
+            phase="validation-start",
+            candidate_head=OLD_HEAD,
+            target_branch="master",
+            target_branch_sha="2" * 40,
+            last_activity=stale,
+            expires_at=stale + timedelta(minutes=120),
+        )
+        return [{
+            "id": 11131,
+            "body": body,
+            "user": {"login": "sysadmin"},
+            "author": "sysadmin",
+            "created_at": stale.isoformat(),
+            "updated_at": stale.isoformat(),
+        }]
+
+    def _assess(self, **kwargs):
+        defaults = dict(
+            pr_number=701,
+            current_head_sha=NEW_HEAD,
+            formal_reviews=[],
+            repo=REPO,
+            expected_repo=REPO,
+            requesting_session_id="fresh-controller",
+            controller_recovery_authorized=True,
+            owner_process_alive=False,
+        )
+        defaults.update(kwargs)
+        return leases.assess_obsolete_reviewer_comment_lease_cleanup(
+            self._comments(), **defaults
+        )
+
+    def test_unknown_worktree_evidence_fails_closed(self):
+        result = self._assess(worktree_exists=None, worktree_clean=None)
+        self.assertEqual(
+            result["classification"], "orphaned_expired_superseded_head"
+        )
+        self.assertFalse(result["cleanup_allowed"])
+        self.assertIn(
+            "affirmative safe evidence",
+            " ".join(result["fail_closed_reasons"]),
+        )
+
+    def test_unknown_cleanliness_with_existing_worktree_fails_closed(self):
+        result = self._assess(worktree_exists=True, worktree_clean=None)
+        self.assertFalse(result["cleanup_allowed"])
+
+    def test_affirmative_clean_worktree_is_eligible(self):
+        result = self._assess(worktree_exists=True, worktree_clean=True)
+        self.assertEqual(
+            result["classification"], "orphaned_expired_superseded_head"
+        )
+        self.assertTrue(result["cleanup_allowed"])
+
+    def test_affirmative_absent_worktree_is_eligible(self):
+        result = self._assess(worktree_exists=False, worktree_clean=None)
+        self.assertTrue(result["cleanup_allowed"])
+
+    def test_matrix_matches_diagnosis_gate(self):
+        # The cleanup matrix and the diagnosis path must agree: unknown
+        # evidence yields acquire-only, affirmative evidence yields cleanup.
+        diagnosis_unknown = leases.diagnose_reviewer_pr_lease_handoff(
+            self._comments(),
+            pr_number=701,
+            current_session_id=None,
+            current_reviewer_identity="fresh-reviewer",
+            current_head_sha=NEW_HEAD,
+            formal_reviews=[],
+            owner_process_alive=False,
+            worktree_exists=None,
+            worktree_clean=None,
+        )
+        assess_unknown = self._assess(worktree_exists=None, worktree_clean=None)
+        self.assertNotEqual(
+            diagnosis_unknown["next_action"],
+            "cleanup_obsolete_reviewer_comment_lease",
+        )
+        self.assertFalse(assess_unknown["cleanup_allowed"])
+
+
+class TestF4CrashShadowDurability(unittest.TestCase):
+    """F4: crash-orphan evidence survives the former 4h TTL boundary."""
+
+    SID = "82984-abcabcabcabc"
+
+    def setUp(self):
+        self._dir = tempfile.TemporaryDirectory()
+        self._env = patch.dict(
+            os.environ, {"GITEA_MCP_SESSION_STATE_DIR": self._dir.name}
+        )
+        self._env.start()
+        leases._SESSION_LEASE = None
+
+    def tearDown(self):
+        self._env.stop()
+        self._dir.cleanup()
+        leases._SESSION_LEASE = None
+
+    def _age_record(self, kind: str, hours: float, instance_id: str | None = None):
+        path = mss.state_file_path(kind=kind, instance_id=instance_id)
+        with open(path, encoding="utf-8") as fh:
+            envelope = json.load(fh)
+        stamp = (
+            (_now() - timedelta(hours=hours)).isoformat().replace("+00:00", "Z")
+        )
+        envelope["recorded_at"] = stamp
+        envelope["updated_at"] = stamp
+        envelope["payload"]["recorded_at"] = stamp
+        envelope["payload"]["updated_at"] = stamp
+        with open(path, "w", encoding="utf-8") as fh:
+            json.dump(envelope, fh)
+
+    def _save_shadow(self) -> None:
+        mss.save_state(
+            kind=mss.KIND_REVIEWER_SESSION_LEASE,
+            instance_id=self.SID,
+            payload={"session_id": self.SID, "owner_pid": os.getpid()},
+        )
+
+    def _load_shadow(self):
+        return mss.load_state(
+            kind=mss.KIND_REVIEWER_SESSION_LEASE, instance_id=self.SID
+        )
+
+    def test_shadow_loads_before_former_ttl_boundary(self):
+        self._save_shadow()
+        self._age_record(
+            mss.KIND_REVIEWER_SESSION_LEASE, hours=3.9, instance_id=self.SID
+        )
+        self.assertIsNotNone(self._load_shadow())
+
+    def test_shadow_loads_at_former_ttl_boundary(self):
+        self._save_shadow()
+        self._age_record(
+            mss.KIND_REVIEWER_SESSION_LEASE, hours=4.0, instance_id=self.SID
+        )
+        self.assertIsNotNone(self._load_shadow())
+
+    def test_shadow_loads_long_after_former_ttl_boundary(self):
+        self._save_shadow()
+        self._age_record(
+            mss.KIND_REVIEWER_SESSION_LEASE, hours=27.0, instance_id=self.SID
+        )
+        record = self._load_shadow()
+        self.assertIsNotNone(record)
+        self.assertEqual(record.get("session_id"), self.SID)
+
+    def test_stale_binding_audit_record_is_also_durable(self):
+        mss.save_state(
+            kind=mss.KIND_STALE_BINDING_RECOVERY,
+            payload={"cleared_env": "GITEA_ACTIVE_WORKTREE"},
+        )
+        self._age_record(mss.KIND_STALE_BINDING_RECOVERY, hours=27.0)
+        self.assertIsNotNone(
+            mss.load_state(kind=mss.KIND_STALE_BINDING_RECOVERY)
+        )
+
+    def test_non_recovery_kinds_keep_ttl(self):
+        mss.save_state(
+            kind=mss.KIND_WORKFLOW_LOAD, payload={"workflow_hash": "abc"}
+        )
+        self._age_record(mss.KIND_WORKFLOW_LOAD, hours=5.0)
+        self.assertIsNone(mss.load_state(kind=mss.KIND_WORKFLOW_LOAD))
+
+    def test_sanctioned_clear_is_the_terminal_reconciliation(self):
+        self._save_shadow()
+        self._age_record(
+            mss.KIND_REVIEWER_SESSION_LEASE, hours=27.0, instance_id=self.SID
+        )
+        mss.clear_state(
+            kind=mss.KIND_REVIEWER_SESSION_LEASE, instance_id=self.SID
+        )
+        self.assertIsNone(self._load_shadow())
+        self.assertEqual(
+            mss.list_states(kind=mss.KIND_REVIEWER_SESSION_LEASE), []
+        )
+
+    def test_crash_orphan_recovery_works_after_former_ttl(self):
+        # End to end: a crashed session's shadow still proves owner exit a
+        # day later.
+        leases.record_session_lease(
+            {
+                "pr_number": 701,
+                "session_id": self.SID,
+                "worktree": "branches/review-pr-654",
+                "candidate_head": OLD_HEAD,
+                "repo": REPO,
+                "comment_id": 11131,
+                "profile": "prgs-reviewer",
+                "reviewer_identity": "sysadmin",
+                "phase": "claimed",
+            },
+            lease_provenance={"source": "acquire", "comment_id": 11131},
+        )
+        leases._SESSION_LEASE = None  # simulated crash: no sanctioned clear
+        self._age_record(
+            mss.KIND_REVIEWER_SESSION_LEASE, hours=27.0, instance_id=self.SID
+        )
+        shadow = leases.load_session_lease_shadow(session_id=self.SID)
+        self.assertIsNotNone(shadow)
+        verdict = leases.assess_crashed_session_lease_orphan(
+            shadow,
+            lease={"session_id": self.SID},
+            current_pid=os.getpid() + 1,
+            pid_alive=False,
+        )
+        self.assertTrue(verdict["orphan_evidence"])
+        self.assertIs(verdict["owner_process_alive"], False)
+
+
+class TestF5ConcurrentSessionShadows(unittest.TestCase):
+    """F5: concurrent same-profile sessions keep distinct shadow records."""
+
+    SID_A = "11111-aaaaaaaaaaaa"
+    SID_B = "22222-bbbbbbbbbbbb"
+
+    def setUp(self):
+        self._dir = tempfile.TemporaryDirectory()
+        self._env = patch.dict(
+            os.environ, {"GITEA_MCP_SESSION_STATE_DIR": self._dir.name}
+        )
+        self._env.start()
+        leases._SESSION_LEASE = None
+
+    def tearDown(self):
+        self._env.stop()
+        self._dir.cleanup()
+        leases._SESSION_LEASE = None
+
+    def _lease(self, session_id: str, pr_number: int, head: str) -> dict:
+        return {
+            "pr_number": pr_number,
+            "session_id": session_id,
+            "worktree": f"branches/review-pr-{pr_number}-{session_id[:5]}",
+            "candidate_head": head,
+            "repo": REPO,
+            "comment_id": 11221,
+            "profile": "prgs-reviewer",
+            "reviewer_identity": "sysadmin",
+            "phase": "claimed",
+        }
+
+    def _record(self, lease: dict) -> None:
+        leases.record_session_lease(
+            lease,
+            lease_provenance={"source": "acquire", "comment_id": 11221},
+        )
+
+    def test_interleaved_sessions_do_not_overwrite_each_other(self):
+        self._record(self._lease(self.SID_A, 703, OLD_HEAD))
+        self._record(self._lease(self.SID_B, 701, NEW_HEAD))
+        # Session A heartbeats again after B wrote.
+        updated_a = self._lease(self.SID_A, 703, OLD_HEAD)
+        updated_a["phase"] = "validation-start"
+        self._record(updated_a)
+
+        shadow_a = leases.load_session_lease_shadow(session_id=self.SID_A)
+        shadow_b = leases.load_session_lease_shadow(session_id=self.SID_B)
+        self.assertEqual(shadow_a.get("session_id"), self.SID_A)
+        self.assertEqual(shadow_a.get("pr_number"), 703)
+        self.assertEqual(shadow_a.get("phase"), "validation-start")
+        self.assertEqual(shadow_b.get("session_id"), self.SID_B)
+        self.assertEqual(shadow_b.get("pr_number"), 701)
+        self.assertEqual(shadow_b.get("candidate_head"), NEW_HEAD)
+
+    def test_shadow_lookup_never_misattributes(self):
+        self._record(self._lease(self.SID_A, 703, OLD_HEAD))
+        self.assertIsNone(
+            leases.load_session_lease_shadow(session_id=self.SID_B)
+        )
+        verdict = leases.assess_crashed_session_lease_orphan(
+            leases.load_session_lease_shadow(session_id=self.SID_A),
+            lease={"session_id": self.SID_B},
+            pid_alive=False,
+        )
+        self.assertFalse(verdict["orphan_evidence"])
+
+    def test_clearing_one_session_preserves_the_other(self):
+        self._record(self._lease(self.SID_A, 703, OLD_HEAD))
+        self._record(self._lease(self.SID_B, 701, NEW_HEAD))
+        # Sanctioned clear of B (the currently recorded in-memory lease).
+        leases.clear_session_lease()
+        self.assertIsNone(
+            leases.load_session_lease_shadow(session_id=self.SID_B)
+        )
+        shadow_a = leases.load_session_lease_shadow(session_id=self.SID_A)
+        self.assertIsNotNone(shadow_a)
+        self.assertEqual(shadow_a.get("session_id"), self.SID_A)
+
+    def test_reconnected_process_recovers_by_session_id(self):
+        self._record(self._lease(self.SID_A, 703, OLD_HEAD))
+        self._record(self._lease(self.SID_B, 701, NEW_HEAD))
+        # Simulated crash + reconnect: in-memory state is gone, durable
+        # records remain enumerable and individually attributable.
+        leases._SESSION_LEASE = None
+        records = mss.list_states(kind=mss.KIND_REVIEWER_SESSION_LEASE)
+        self.assertEqual(
+            sorted(r.get("session_id") for r in records),
+            sorted([self.SID_A, self.SID_B]),
+        )
+        shadow = leases.load_session_lease_shadow(session_id=self.SID_B)
+        self.assertEqual(shadow.get("pr_number"), 701)
+
+    def test_legacy_single_slot_record_still_resolves_by_session_id(self):
+        # Upgrade path: a record written by pre-F5 code (no instance key)
+        # remains usable when its payload names the requested session.
+        mss.save_state(
+            kind=mss.KIND_REVIEWER_SESSION_LEASE,
+            payload={"session_id": self.SID_A, "owner_pid": os.getpid()},
+        )
+        shadow = leases.load_session_lease_shadow(session_id=self.SID_A)
+        self.assertIsNotNone(shadow)
+        self.assertIsNone(
+            leases.load_session_lease_shadow(session_id=self.SID_B)
+        )
+
+
+class TestF6AuditBeforeClear(_CapabilityHarness):
+    """F6: the durable audit record is persisted before the env clears."""
+
+    def _recovery_env(self, missing: str) -> dict:
+        return self._env(
+            GITEA_ACTIVE_WORKTREE=missing,
+            GITEA_FORCE_STALE_BINDING_RECOVERY="1",
+        )
+
+    def test_audit_write_failure_blocks_the_clear(self):
+        missing = self._deleted_worktree()
+        with patch.dict(os.environ, self._recovery_env(missing)):
+            with patch.object(
+                mss, "save_state", side_effect=OSError("disk full")
+            ):
+                report = mcp_server._assess_stale_active_binding(
+                    auto_recover=True
+                )
+            self.assertEqual(os.environ.get("GITEA_ACTIVE_WORKTREE"), missing)
+        self.assertFalse(report["recovery_performed"])
+        self.assertIs(report.get("audit_persisted"), False)
+        self.assertTrue(report.get("audit_write_failed"))
+        self.assertIn("NOT cleared", " ".join(report.get("reasons") or []))
+
+    def test_retry_after_audit_failure_recovers(self):
+        missing = self._deleted_worktree()
+        with patch.dict(os.environ, self._recovery_env(missing)):
+            with patch.object(
+                mss, "save_state", side_effect=OSError("disk full")
+            ):
+                first = mcp_server._assess_stale_active_binding(
+                    auto_recover=True
+                )
+            self.assertFalse(first["recovery_performed"])
+            # Persistence recovered; the retry clears with a durable audit.
+            second = mcp_server._assess_stale_active_binding(auto_recover=True)
+            self.assertTrue(second["recovery_performed"])
+            self.assertIs(second.get("audit_persisted"), True)
+            self.assertNotIn("GITEA_ACTIVE_WORKTREE", os.environ)
+            audit = mss.load_state(kind=mss.KIND_STALE_BINDING_RECOVERY)
+        self.assertIsNotNone(audit)
+        self.assertEqual(audit.get("recovery_status"), "performed")
+        self.assertEqual(audit.get("cleared_value"), missing)
+
+    def test_audit_record_exists_before_env_clear(self):
+        missing = self._deleted_worktree()
+        observed: list[tuple[str | None, str | None]] = []
+        real_save = mss.save_state
+
+        def _spy(**kwargs):
+            observed.append(
+                (
+                    (kwargs.get("payload") or {}).get("recovery_status"),
+                    os.environ.get("GITEA_ACTIVE_WORKTREE"),
+                )
+            )
+            return real_save(**kwargs)
+
+        with patch.dict(os.environ, self._recovery_env(missing)):
+            with patch.object(mss, "save_state", side_effect=_spy):
+                report = mcp_server._assess_stale_active_binding(
+                    auto_recover=True
+                )
+        self.assertTrue(report["recovery_performed"])
+        self.assertGreaterEqual(len(observed), 2)
+        pending_status, env_at_pending = observed[0]
+        performed_status, env_at_performed = observed[-1]
+        # Ordering proof: the binding was still present while the pending
+        # audit persisted, and gone by the time the performed status wrote.
+        self.assertEqual(pending_status, "pending")
+        self.assertEqual(env_at_pending, missing)
+        self.assertEqual(performed_status, "performed")
+        self.assertIsNone(env_at_performed)
+
+    def test_recovery_is_idempotent_after_success(self):
+        missing = self._deleted_worktree()
+        with patch.dict(os.environ, self._recovery_env(missing)):
+            first = mcp_server._assess_stale_active_binding(auto_recover=True)
+            second = mcp_server._assess_stale_active_binding(auto_recover=True)
+        self.assertTrue(first["recovery_performed"])
+        self.assertEqual(second["classification"], sbr.CLASSIFICATION_UNBOUND)
+        self.assertFalse(second["recovery_performed"])
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_issue_702_stale_binding_lease_recovery.py

@@ -0,0 +1,517 @@
+"""Regression tests for #702: stale runtime binding + orphaned durable lease.
+
+Reproduces the PR #701 incident (lease 65664-4f248d213d60, comments
+11129/11131): the MCP daemon crashed without teardown, destroying the
+in-memory session lease while the durable comment lease survived, and the
+auto-reconnected daemon inherited a stale ``GITEA_ACTIVE_WORKTREE`` binding
+(``branches/review-pr-654``) from its parent environment.
+
+Covers the five mandated scenarios:
+
+1. lost in-session leases  — durable shadow survives a crash and yields
+   provable owner-process evidence
+2. old-head leases         — expired superseded-head lease with no terminal
+   review becomes recoverable only with orphan evidence + controller authority
+3. runtime/worktree mismatch — env bindings to deleted worktrees are demoted,
+   never silently bound
+4. managed reconnect       — boot-inherited uncorroborated bindings are
+   surfaced for managed recovery; provably stale ones are clear-eligible
+5. safe expiry             — pre-expiry stays fail-closed wait (no steal);
+   canonical expiry unlocks acquisition and guarded cleanup
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+import unittest
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+import namespace_workspace_binding as nwb  # noqa: E402
+import reviewer_pr_lease as leases  # noqa: E402
+import stale_binding_recovery as sbr  # noqa: E402
+
+# PR #701 incident fixtures.
+OLD_HEAD = "b4d0cb22e452a07c8e816745c704ddb0f43edf0b"
+NEW_HEAD = "6b675f5c834b41f9d74e8a54294ff44dddf28ae4"
+CRASHED_SESSION = "65664-4f248d213d60"
+LEASE_COMMENT = 11131
+REPO = "Scaled-Tech-Consulting/Gitea-Tools"
+PR = 701
+ISSUE = 699
+LEASE_WORKTREE = "branches/review-fix-issue-699-structured-auth-mcp-errors"
+
+
+def _now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+def _lease_comment(
+    *,
+    phase: str = "validation-start",
+    candidate_head: str = OLD_HEAD,
+    session_id: str = CRASHED_SESSION,
+    comment_id: int = LEASE_COMMENT,
+    last_activity: datetime | None = None,
+    expires_at: datetime | None = None,
+    worktree: str = LEASE_WORKTREE,
+) -> dict:
+    stamp = last_activity or _now()
+    body = leases.format_lease_body(
+        repo=REPO,
+        pr_number=PR,
+        issue_number=ISSUE,
+        reviewer_identity="sysadmin",
+        profile="prgs-reviewer",
+        session_id=session_id,
+        worktree=worktree,
+        phase=phase,
+        candidate_head=candidate_head,
+        target_branch="master",
+        target_branch_sha="2" * 40,
+        last_activity=stamp,
+        expires_at=expires_at,
+    )
+    return {
+        "id": comment_id,
+        "body": body,
+        "user": {"login": "sysadmin"},
+        "author": "sysadmin",
+        "created_at": stamp.isoformat(),
+        "updated_at": stamp.isoformat(),
+    }
+
+
+def _expired_superseded_comments() -> list[dict]:
+    stale = _now() - timedelta(hours=3)
+    return [_lease_comment(
+        last_activity=stale,
+        expires_at=stale + timedelta(minutes=120),
+    )]
+
+
+def _fresh_superseded_comments() -> list[dict]:
+    recent = _now() - timedelta(minutes=10)
+    return [_lease_comment(
+        last_activity=recent,
+        expires_at=recent + timedelta(minutes=120),
+    )]
+
+
+class TestLostInSessionLeaseShadow(unittest.TestCase):
+    """Scenario 1: the durable shadow survives a daemon crash (#702 AC3)."""
+
+    def setUp(self):
+        leases._SESSION_LEASE = None
+
+    def tearDown(self):
+        leases._SESSION_LEASE = None
+
+    def _record(self) -> dict:
+        return leases.record_session_lease(
+            {
+                "pr_number": PR,
+                "issue_number": ISSUE,
+                "session_id": CRASHED_SESSION,
+                "reviewer_identity": "sysadmin",
+                "profile": "prgs-reviewer",
+                "worktree": LEASE_WORKTREE,
+                "phase": "claimed",
+                "candidate_head": OLD_HEAD,
+                "repo": REPO,
+                "comment_id": LEASE_COMMENT,
+            },
+            lease_provenance={"source": "acquire", "comment_id": LEASE_COMMENT},
+        )
+
+    def test_shadow_written_on_record_and_survives_simulated_crash(self):
+        self._record()
+        # Simulated unhandled crash: the in-memory dict dies with the process
+        # but the sanctioned clear path never runs.
+        leases._SESSION_LEASE = None
+        shadow = leases.load_session_lease_shadow()
+        self.assertIsNotNone(shadow)
+        self.assertEqual(shadow.get("session_id"), CRASHED_SESSION)
+        self.assertEqual(shadow.get("pr_number"), PR)
+        self.assertEqual(int(shadow.get("owner_pid")), os.getpid())
+
+    def test_sanctioned_clear_removes_shadow(self):
+        self._record()
+        leases.clear_session_lease()
+        self.assertIsNone(leases.load_session_lease_shadow())
+
+    def test_orphan_assessment_dead_owner_pid_proves_exit(self):
+        self._record()
+        leases._SESSION_LEASE = None
+        shadow = leases.load_session_lease_shadow()
+        lease = {"session_id": CRASHED_SESSION}
+        verdict = leases.assess_crashed_session_lease_orphan(
+            shadow, lease=lease, current_pid=os.getpid() + 1, pid_alive=False
+        )
+        self.assertTrue(verdict["orphan_evidence"])
+        self.assertIs(verdict["owner_process_alive"], False)
+
+    def test_orphan_assessment_alive_pid_is_not_ownership_proof(self):
+        self._record()
+        shadow = leases.load_session_lease_shadow()
+        verdict = leases.assess_crashed_session_lease_orphan(
+            shadow,
+            lease={"session_id": CRASHED_SESSION},
+            current_pid=os.getpid() + 1,
+            pid_alive=True,
+        )
+        self.assertFalse(verdict["orphan_evidence"])
+        self.assertIsNone(verdict["owner_process_alive"])
+
+    def test_orphan_assessment_session_mismatch_proves_nothing(self):
+        self._record()
+        shadow = leases.load_session_lease_shadow()
+        verdict = leases.assess_crashed_session_lease_orphan(
+            shadow, lease={"session_id": "other-999"}, pid_alive=False
+        )
+        self.assertFalse(verdict["orphan_evidence"])
+        self.assertIsNone(verdict["owner_process_alive"])
+
+    def test_orphan_assessment_without_shadow_is_unknown(self):
+        verdict = leases.assess_crashed_session_lease_orphan(
+            None, lease={"session_id": CRASHED_SESSION}
+        )
+        self.assertFalse(verdict["orphan_evidence"])
+        self.assertIsNone(verdict["owner_process_alive"])
+
+
+class TestOldHeadLeaseCleanup(unittest.TestCase):
+    """Scenario 2: expired superseded-head lease with no terminal review."""
+
+    def _assess(self, comments, **kwargs):
+        defaults = dict(
+            pr_number=PR,
+            current_head_sha=NEW_HEAD,
+            formal_reviews=[],
+            repo=REPO,
+            expected_repo=REPO,
+            requesting_session_id="fresh-controller",
+            controller_recovery_authorized=True,
+            worktree_exists=True,
+            worktree_clean=True,
+            owner_process_alive=False,
+        )
+        defaults.update(kwargs)
+        return leases.assess_obsolete_reviewer_comment_lease_cleanup(
+            comments, **defaults
+        )
+
+    def test_expired_orphan_with_full_evidence_is_cleanup_eligible(self):
+        result = self._assess(_expired_superseded_comments())
+        self.assertEqual(result["classification"], "orphaned_expired_superseded_head")
+        self.assertTrue(result["cleanup_allowed"])
+        self.assertIn("phase: released", result["release_body"] or "")
+        self.assertEqual(
+            result["exact_next_action"], "cleanup_obsolete_reviewer_comment_lease"
+        )
+
+    def test_expired_orphan_without_owner_evidence_fails_closed(self):
+        result = self._assess(
+            _expired_superseded_comments(), owner_process_alive=None
+        )
+        self.assertFalse(result["cleanup_allowed"])
+        self.assertIn(
+            "provable owner-process-exit evidence",
+            " ".join(result["fail_closed_reasons"]),
+        )
+
+    def test_expired_orphan_without_controller_authority_fails_closed(self):
+        result = self._assess(
+            _expired_superseded_comments(), controller_recovery_authorized=False
+        )
+        self.assertFalse(result["cleanup_allowed"])
+
+    def test_expired_orphan_with_dirty_worktree_fails_closed(self):
+        result = self._assess(
+            _expired_superseded_comments(),
+            worktree_clean=False,
+            worktree_has_unpreserved_work=True,
+        )
+        self.assertFalse(result["cleanup_allowed"])
+
+    def test_unexpired_superseded_no_terminal_stays_ambiguous_wait(self):
+        # Regression guard: pre-expiry there is still no steal path.
+        result = self._assess(_fresh_superseded_comments())
+        self.assertEqual(result["classification"], "ambiguous_conflicting_evidence")
+        self.assertFalse(result["cleanup_allowed"])
+        self.assertEqual(result["exact_next_action"], "wait")
+
+
+class TestRuntimeWorktreeMismatch(unittest.TestCase):
+    """Scenario 3: env bindings to deleted worktrees are demoted (#702 AC2)."""
+
+    def test_missing_active_env_binding_is_demoted(self):
+        demotions: list[str] = []
+        missing = "/nonexistent/branches/review-pr-654"
+        path, source = nwb.resolve_namespace_workspace(
+            role_kind="reviewer",
+            process_project_root=os.getcwd(),
+            env={"GITEA_ACTIVE_WORKTREE": missing},
+            demotions=demotions,
+            verify_paths=True,
+        )
+        self.assertEqual(source, "MCP server process root (default)")
+        self.assertEqual(len(demotions), 1)
+        self.assertIn("no longer exists", demotions[0])
+
+    def test_missing_active_falls_through_to_existing_role_env(self):
+        existing = os.getcwd()
+        path, source = nwb.resolve_namespace_workspace(
+            role_kind="reviewer",
+            process_project_root="/tmp",
+            env={
+                "GITEA_ACTIVE_WORKTREE": "/nonexistent/branches/review-pr-654",
+                "GITEA_REVIEWER_WORKTREE": existing,
+            },
+            verify_paths=True,
+        )
+        self.assertEqual(path, os.path.realpath(existing))
+        self.assertEqual(source, "GITEA_REVIEWER_WORKTREE environment variable")
+
+    def test_existing_active_env_binding_still_wins(self):
+        existing = os.getcwd()
+        path, source = nwb.resolve_namespace_workspace(
+            role_kind="reviewer",
+            process_project_root="/tmp",
+            env={"GITEA_ACTIVE_WORKTREE": existing},
+            verify_paths=True,
+        )
+        self.assertEqual(path, os.path.realpath(existing))
+        self.assertEqual(source, "GITEA_ACTIVE_WORKTREE environment variable")
+
+    def test_explicit_worktree_path_argument_is_never_demoted(self):
+        missing = "/nonexistent/branches/explicit"
+        path, source = nwb.resolve_namespace_workspace(
+            role_kind="reviewer",
+            worktree_path=missing,
+            process_project_root="/tmp",
+            env={},
+            verify_paths=True,
+        )
+        self.assertEqual(source, "worktree_path argument")
+
+    def test_mutation_context_reports_demotion_in_ignored_bindings(self):
+        ctx = nwb.resolve_namespace_mutation_context(
+            role_kind="reviewer",
+            worktree_path=None,
+            process_project_root=os.getcwd(),
+            env={"GITEA_ACTIVE_WORKTREE": "/nonexistent/branches/review-pr-654"},
+        )
+        joined = " ".join(ctx["ignored_bindings"])
+        self.assertIn("stale binding, #702", joined)
+
+
+class TestManagedReconnectRecovery(unittest.TestCase):
+    """Scenario 4: boot-inherited bindings and the sanctioned recovery plan."""
+
+    def test_uncorroborated_inherited_reviewer_binding_is_unverified(self):
+        classification = sbr.classify_active_worktree_binding(
+            active_value=os.getcwd(),
+            role_env_value=None,
+            session_lease_worktree=None,
+            boot_inherited=True,
+            path_exists=True,
+            role_kind="reviewer",
+        )
+        self.assertEqual(
+            classification["classification"], sbr.CLASSIFICATION_UNVERIFIED_INHERITED
+        )
+        self.assertFalse(classification["clear_eligible"])
+        plan = sbr.plan_recovery(classification)
+        self.assertFalse(plan["clear_allowed"])
+        self.assertIn("managed", plan["exact_next_action"])
+        self.assertIn("do not clear or rewrite", plan["exact_next_action"])
+
+    def test_missing_path_binding_is_provably_stale_and_clear_eligible(self):
+        classification = sbr.classify_active_worktree_binding(
+            active_value="/nonexistent/branches/review-pr-654",
+            boot_inherited=True,
+            path_exists=False,
+            role_kind="reviewer",
+        )
+        self.assertEqual(
+            classification["classification"],
+            sbr.CLASSIFICATION_PROVABLY_STALE_MISSING_PATH,
+        )
+        plan = sbr.plan_recovery(classification)
+        self.assertTrue(plan["clear_allowed"])
+
+    def test_inherited_binding_superseded_by_session_lease_is_clear_eligible(self):
+        classification = sbr.classify_active_worktree_binding(
+            active_value="/tmp",
+            session_lease_worktree=os.getcwd(),
+            boot_inherited=True,
+            path_exists=True,
+            role_kind="reviewer",
+        )
+        self.assertEqual(
+            classification["classification"],
+            sbr.CLASSIFICATION_SUPERSEDED_BY_SESSION_LEASE,
+        )
+        self.assertTrue(sbr.plan_recovery(classification)["clear_allowed"])
+
+    def test_post_boot_binding_conflict_is_surfaced_not_cleared(self):
+        classification = sbr.classify_active_worktree_binding(
+            active_value="/tmp",
+            session_lease_worktree=os.getcwd(),
+            boot_inherited=False,
+            path_exists=True,
+            role_kind="reviewer",
+        )
+        self.assertEqual(
+            classification["classification"], sbr.CLASSIFICATION_UNVERIFIED_INHERITED
+        )
+        self.assertFalse(sbr.plan_recovery(classification)["clear_allowed"])
+
+    def test_corroborated_bindings_are_untouched(self):
+        for kwargs in (
+            dict(role_env_value=os.getcwd()),
+            dict(session_lease_worktree=os.getcwd()),
+        ):
+            classification = sbr.classify_active_worktree_binding(
+                active_value=os.getcwd(),
+                boot_inherited=True,
+                path_exists=True,
+                role_kind="reviewer",
+                **kwargs,
+            )
+            self.assertEqual(
+                classification["classification"], sbr.CLASSIFICATION_CORROBORATED
+            )
+            self.assertFalse(sbr.plan_recovery(classification)["clear_allowed"])
+
+    def test_author_inherited_binding_stays_corroborated(self):
+        classification = sbr.classify_active_worktree_binding(
+            active_value=os.getcwd(),
+            boot_inherited=True,
+            path_exists=True,
+            role_kind="author",
+        )
+        self.assertEqual(
+            classification["classification"], sbr.CLASSIFICATION_CORROBORATED
+        )
+
+    def test_apply_recovery_clears_env_only_for_authorized_plan(self):
+        env = {"GITEA_ACTIVE_WORKTREE": "/nonexistent/branches/review-pr-654"}
+        plan = sbr.plan_recovery(
+            sbr.classify_active_worktree_binding(
+                active_value=env["GITEA_ACTIVE_WORKTREE"],
+                boot_inherited=True,
+                path_exists=False,
+                role_kind="reviewer",
+            )
+        )
+        applied = sbr.apply_recovery(plan, env=env)
+        self.assertTrue(applied["performed"])
+        self.assertNotIn("GITEA_ACTIVE_WORKTREE", env)
+        self.assertIn("review-pr-654", applied["cleared_value"])
+
+    def test_apply_recovery_refuses_unauthorized_plan(self):
+        env = {"GITEA_ACTIVE_WORKTREE": os.getcwd()}
+        plan = sbr.plan_recovery(
+            sbr.classify_active_worktree_binding(
+                active_value=env["GITEA_ACTIVE_WORKTREE"],
+                boot_inherited=True,
+                path_exists=True,
+                role_kind="reviewer",
+            )
+        )
+        applied = sbr.apply_recovery(plan, env=env)
+        self.assertFalse(applied["performed"])
+        self.assertIn("GITEA_ACTIVE_WORKTREE", env)
+
+
+class TestSafeExpiryDiagnosis(unittest.TestCase):
+    """Scenario 5: canonical expiry flips wait into acquire/guarded cleanup."""
+
+    def _diagnose(self, comments, **kwargs):
+        defaults = dict(
+            pr_number=PR,
+            current_session_id=None,
+            current_reviewer_identity="fresh-reviewer",
+            current_head_sha=NEW_HEAD,
+            formal_reviews=[],
+        )
+        defaults.update(kwargs)
+        return leases.diagnose_reviewer_pr_lease_handoff(comments, **defaults)
+
+    def setUp(self):
+        leases._SESSION_LEASE = None
+
+    def test_pre_expiry_superseded_no_terminal_stays_wait(self):
+        diagnosis = self._diagnose(_fresh_superseded_comments())
+        self.assertEqual(
+            diagnosis["classification"], "ambiguous_conflicting_evidence"
+        )
+        self.assertEqual(diagnosis["next_action"], "wait")
+
+    def test_post_expiry_without_orphan_evidence_unlocks_acquire(self):
+        diagnosis = self._diagnose(_expired_superseded_comments())
+        self.assertEqual(
+            diagnosis["classification"], "orphaned_expired_superseded_head"
+        )
+        self.assertEqual(diagnosis["next_action"], "acquire")
+
+    def test_post_expiry_with_orphan_evidence_offers_guarded_cleanup(self):
+        diagnosis = self._diagnose(
+            _expired_superseded_comments(),
+            owner_process_alive=False,
+            worktree_clean=True,
+            worktree_exists=True,
+        )
+        self.assertEqual(
+            diagnosis["classification"], "orphaned_expired_superseded_head"
+        )
+        self.assertEqual(
+            diagnosis["next_action"], "cleanup_obsolete_reviewer_comment_lease"
+        )
+
+    def test_expired_lease_no_longer_blocks_fresh_acquisition(self):
+        assessment = leases.assess_acquire_lease(
+            _expired_superseded_comments(),
+            pr_number=PR,
+            reviewer_identity="fresh-reviewer",
+            profile="prgs-reviewer",
+            session_id="99999-fresh",
+            repo=REPO,
+            issue_number=ISSUE,
+            worktree="branches/review-pr-701-fresh",
+            candidate_head=NEW_HEAD,
+            target_branch="master",
+            target_branch_sha="2" * 40,
+        )
+        self.assertTrue(assessment.get("acquire_allowed"))
+
+    def test_unparseable_expiry_fails_closed_in_cleanup(self):
+        comment = _lease_comment(
+            last_activity=_now() - timedelta(hours=3), expires_at=None
+        )
+        comment["body"] = comment["body"].replace(
+            "expires_at: ", "expires_at: not-a-timestamp"
+        )
+        result = leases.assess_obsolete_reviewer_comment_lease_cleanup(
+            [comment],
+            pr_number=PR,
+            current_head_sha=NEW_HEAD,
+            formal_reviews=[],
+            repo=REPO,
+            expected_repo=REPO,
+            requesting_session_id="fresh-controller",
+            controller_recovery_authorized=True,
+            worktree_exists=True,
+            worktree_clean=True,
+            owner_process_alive=False,
+        )
+        self.assertFalse(result["cleanup_allowed"])
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_issue_720_expired_decision_lock.py

@@ -0,0 +1,499 @@
+"""#720: Expired old-head KIND_DECISION_LOCK must not block fresh review.
+
+Reproduction shape (PR #616):
+  * REQUEST_CHANGES terminal at head A
+  * open PR advanced to head B
+  * durable decision lock age > default 4h TTL
+  * fresh_review_on_current_head_allowed is true but unreachable under TTL-first reject
+  * mark_final fails with "session state expired after 4h"
+  * assessment may report "no lock" while the file remains on disk
+
+Security invariants preserved:
+  * same-head second terminal remains fail-closed (#332/#620)
+  * REQUEST_CHANGES is never treated as approval
+  * historical mutations remain on the ledger
+  * merged/closed moot cleanup still allowed (#594)
+  * irrecoverable provenance still requires #709 authorization
+  * ordinary profiles do not gain gitea.decision_lock.irrecoverable_recovery
+"""
+
+from __future__ import annotations
+
+import os
+import tempfile
+import unittest
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+from unittest.mock import patch
+
+import sys
+
+ROOT = str(Path(__file__).resolve().parent.parent)
+if ROOT not in sys.path:
+    sys.path.insert(0, ROOT)
+
+import mcp_session_state as ss
+import mcp_server
+import stale_review_decision_lock as srdl
+import task_capability_map
+
+HEAD_A = "a0fffae576673ba7df7456b32e6aec916581bdfb"
+HEAD_B = "a6a2243aad9c3e385fc70509f4941a0f0ec33162"
+HEAD_SAME = HEAD_A
+
+RC_616_A = {
+    "pr_number": 616,
+    "action": "request_changes",
+    "review_id": 443,
+    "review_state": "request_changes",
+    "head_sha": HEAD_A,
+}
+
+
+def _lock(mutations=None, **kwargs):
+    base = {
+        "task": "review_pr",
+        "kind": ss.KIND_DECISION_LOCK,
+        "remote": "prgs",
+        "org": "Scaled-Tech-Consulting",
+        "repo": "Gitea-Tools",
+        "session_pid": os.getpid(),
+        "session_profile": "prgs-reviewer",
+        "session_profile_lock": "prgs-reviewer",
+        "profile_identity": "prgs-reviewer",
+        "final_review_decision_ready": False,
+        "ready_pr_number": kwargs.get("ready_pr"),
+        "ready_action": kwargs.get("ready_action"),
+        "ready_expected_head_sha": kwargs.get("ready_head"),
+        "ready_remote": "prgs" if kwargs.get("ready_pr") else None,
+        "ready_org": "Scaled-Tech-Consulting" if kwargs.get("ready_pr") else None,
+        "ready_repo": "Gitea-Tools" if kwargs.get("ready_pr") else None,
+        "live_mutations": list(mutations or []),
+        "correction_authorized": False,
+        "correction_reason": None,
+    }
+    return base
+
+
+def _open_pr(pr_number=616, head=HEAD_B, merged=False, closed=False):
+    state = "closed" if closed or merged else "open"
+    return {
+        "number": pr_number,
+        "state": state,
+        "merged": merged,
+        "merged_at": "2026-07-16T12:00:00Z" if merged else None,
+        "merge_commit_sha": "m" * 40 if merged else None,
+        "head": {"sha": head},
+    }
+
+
+def _no_lease():
+    return {"block": False, "reasons": [], "mutation_allowed": True}
+
+
+def _feedback(blocking=False, stale=True):
+    return {
+        "success": True,
+        "has_blocking_change_requests": blocking,
+        "review_feedback_stale": stale,
+        "current_head_sha": HEAD_B,
+    }
+
+
+def _age_lock_payload(lock: dict, hours: float = 5.0) -> dict:
+    """Rewrite timestamps to *hours* ago (simulates durable age without touching prod)."""
+    aged = (datetime.now(timezone.utc) - timedelta(hours=hours)).isoformat().replace(
+        "+00:00", "Z"
+    )
+    out = dict(lock)
+    out["recorded_at"] = aged
+    out["updated_at"] = aged
+    return out
+
+
+class TestIssue720ExpiredDecisionLockLifecycle(unittest.TestCase):
+    def setUp(self):
+        self._tmp = tempfile.TemporaryDirectory()
+        self.env = patch.dict(
+            os.environ,
+            {
+                ss.STATE_DIR_ENV: self._tmp.name,
+                ss.SESSION_PROFILE_LOCK_ENV: "prgs-reviewer",
+                "GITEA_MCP_PROFILE": "prgs-reviewer",
+                "GITEA_PROFILE_NAME": "prgs-reviewer",
+            },
+            clear=False,
+        )
+        self.env.start()
+        mcp_server._REVIEW_DECISION_LOCK = None
+        import review_workflow_load
+
+        review_workflow_load.clear_review_workflow_load()
+        review_workflow_load.record_review_workflow_load(mcp_server.PROJECT_ROOT)
+        mcp_server.gitea_load_review_workflow()
+
+    def tearDown(self):
+        mcp_server._REVIEW_DECISION_LOCK = None
+        import review_workflow_load
+
+        review_workflow_load.clear_review_workflow_load()
+        self.env.stop()
+        self._tmp.cleanup()
+
+    def _persist_aged_lock(self, mutations, hours: float = 5.0, **kwargs):
+        """Write decision lock aged > TTL to the temp durable store (test-only)."""
+        payload = _age_lock_payload(_lock(mutations, **kwargs), hours=hours)
+        saved = ss.save_state(
+            kind=ss.KIND_DECISION_LOCK,
+            payload=payload,
+            remote="prgs",
+            org="Scaled-Tech-Consulting",
+            repo="Gitea-Tools",
+            profile_identity="prgs-reviewer",
+        )
+        # save_state refreshes updated_at; re-age the on-disk envelope for TTL tests.
+        path = ss.state_file_path(
+            kind=ss.KIND_DECISION_LOCK,
+            profile_identity="prgs-reviewer",
+            state_dir=self._tmp.name,
+        )
+        aged = (datetime.now(timezone.utc) - timedelta(hours=hours)).isoformat().replace(
+            "+00:00", "Z"
+        )
+        import json
+
+        with open(path, encoding="utf-8") as fh:
+            envelope = json.load(fh)
+        envelope["recorded_at"] = aged
+        envelope["updated_at"] = aged
+        if isinstance(envelope.get("payload"), dict):
+            envelope["payload"]["recorded_at"] = aged
+            envelope["payload"]["updated_at"] = aged
+        with open(path, "w", encoding="utf-8") as fh:
+            json.dump(envelope, fh, indent=2, sort_keys=True)
+            fh.write("\n")
+        # Drop memory so subsequent loads hit durable store.
+        mcp_server._REVIEW_DECISION_LOCK = None
+        return saved
+
+    def _mark(self, pr, action, head, feedback=None):
+        with patch("mcp_server._list_pr_lease_comments", return_value=[]), patch(
+            "mcp_server._pr_work_lease_reviewer_block", return_value=_no_lease()
+        ), patch.object(
+            mcp_server,
+            "gitea_get_pr_review_feedback",
+            return_value=feedback or _feedback(blocking=False, stale=True),
+        ), patch.object(
+            mcp_server,
+            "gitea_check_pr_eligibility",
+            return_value={"eligible": True, "head_sha": head},
+        ), patch.object(
+            mcp_server.mcp_daemon_guard,
+            "assert_sanctioned_mutation_runtime",
+            return_value=None,
+        ), patch.object(
+            mcp_server.mcp_daemon_guard,
+            "assert_no_direct_import_bypass",
+            return_value=None,
+        ):
+            return mcp_server.gitea_mark_final_review_decision(
+                pr_number=pr,
+                action=action,
+                expected_head_sha=head,
+                remote="prgs",
+                org="Scaled-Tech-Consulting",
+                repo="Gitea-Tools",
+            )
+
+    # --- AC regression matrix ---
+
+    def test_under_four_hours_fresh_review_at_b_allowed(self):
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=1.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        res = self._mark(616, "approve", HEAD_B)
+        self.assertTrue(res.get("marked_ready"), res)
+
+    def test_over_four_hours_fresh_review_at_b_still_allowed(self):
+        """Primary #720 defect: age > TTL must not block head-B mark_final."""
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=5.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        # Prove durable load is possible for recovery-critical decision locks.
+        loaded = ss.load_state(
+            kind=ss.KIND_DECISION_LOCK,
+            profile_identity="prgs-reviewer",
+        )
+        self.assertIsNotNone(
+            loaded,
+            "expired KIND_DECISION_LOCK must remain loadable (not generic TTL cache)",
+        )
+        res = self._mark(616, "approve", HEAD_B)
+        self.assertTrue(
+            res.get("marked_ready"),
+            f"expected mark_ready on head B after >4h; got {res}",
+        )
+        reasons = " ".join(res.get("reasons") or [])
+        self.assertNotIn("session state expired", reasons)
+
+    def test_historical_review_at_a_preserved_and_stale(self):
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=5.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        res = self._mark(616, "approve", HEAD_B)
+        self.assertTrue(res.get("marked_ready"), res)
+        lock = mcp_server._load_review_decision_lock()
+        self.assertIsNotNone(lock)
+        hist = [m for m in lock["live_mutations"] if m.get("review_id") == 443]
+        self.assertEqual(len(hist), 1)
+        self.assertEqual(hist[0]["head_sha"], HEAD_A)
+        self.assertEqual(hist[0]["action"], "request_changes")
+        a = srdl.assess_stale_review_decision_lock(
+            lock, pr_live=_open_pr(616, HEAD_B)
+        )
+        self.assertTrue(a["stale_by_head"])
+        self.assertTrue(a["fresh_review_on_current_head_allowed"])
+        self.assertEqual(a["locked_head_sha"], HEAD_A)
+
+    def test_no_reuse_approval_from_head_a(self):
+        """REQUEST_CHANGES at A is never an approval credential for B."""
+        self._persist_aged_lock([RC_616_A], hours=5.0)
+        lock = ss.load_state(
+            kind=ss.KIND_DECISION_LOCK, profile_identity="prgs-reviewer"
+        )
+        last = srdl.last_terminal_mutation(lock)
+        self.assertEqual(last.get("action"), "request_changes")
+        self.assertNotEqual(last.get("action"), "approve")
+        # Gate must still require a new mark/submit for head B; prior RC is not approve.
+        reasons = mcp_server.terminal_review_hard_stop_reasons(
+            616, "mark_ready", expected_head_sha=HEAD_B
+        )
+        self.assertEqual(reasons, [])
+
+    def test_second_terminal_mutation_at_b_same_run_blocked(self):
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=5.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        res = self._mark(616, "approve", HEAD_B)
+        self.assertTrue(res.get("marked_ready"), res)
+        # Record terminal approve at B (same run).
+        lock = mcp_server._load_review_decision_lock()
+        lock["final_review_decision_ready"] = True
+        lock["ready_pr_number"] = 616
+        lock["ready_action"] = "approve"
+        lock["ready_expected_head_sha"] = HEAD_B
+        mcp_server._save_review_decision_lock(lock)
+        mcp_server.record_live_review_mutation(616, "approve", review_id=999)
+        # Second terminal at same head B must hard-stop.
+        hard = mcp_server.terminal_review_hard_stop_reasons(
+            616, "mark_ready", expected_head_sha=HEAD_B
+        )
+        self.assertTrue(hard)
+        res2 = self._mark(616, "approve", HEAD_B)
+        self.assertFalse(res2.get("marked_ready"))
+
+    def test_open_pr_same_head_expired_still_fail_closed(self):
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=5.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        res = self._mark(616, "approve", HEAD_SAME)
+        self.assertFalse(res.get("marked_ready"), res)
+        self.assertTrue(
+            any("#332" in r or "already consumed" in r for r in (res.get("reasons") or [])),
+            res,
+        )
+
+    def test_merged_pr_moot_cleanup_still_allowed(self):
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=5.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        lock = ss.load_state(
+            kind=ss.KIND_DECISION_LOCK, profile_identity="prgs-reviewer"
+        )
+        a = srdl.assess_stale_review_decision_lock(
+            lock, pr_live=_open_pr(616, HEAD_A, merged=True)
+        )
+        self.assertTrue(a["is_moot"])
+        self.assertTrue(a["cleanup_allowed"])
+        self.assertFalse(a["fresh_review_on_current_head_allowed"])
+
+    def test_assessment_reports_expired_old_head_not_absent(self):
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=5.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        # Disk presence + load after lifecycle fix.
+        path = ss.state_file_path(
+            kind=ss.KIND_DECISION_LOCK,
+            profile_identity="prgs-reviewer",
+            state_dir=self._tmp.name,
+        )
+        self.assertTrue(os.path.exists(path))
+        loaded = ss.load_state(
+            kind=ss.KIND_DECISION_LOCK, profile_identity="prgs-reviewer"
+        )
+        self.assertIsNotNone(loaded)
+        inspect = ss.inspect_state_envelope(
+            kind=ss.KIND_DECISION_LOCK,
+            profile_identity="prgs-reviewer",
+            state_dir=self._tmp.name,
+        )
+        self.assertTrue(inspect.get("on_disk"))
+        self.assertTrue(inspect.get("has_payload"))
+        self.assertTrue(inspect.get("recovery_critical") or inspect.get("ttl_exempt"))
+        self.assertTrue(inspect.get("age_hours", 0) >= 4.0)
+        a = srdl.assess_stale_review_decision_lock(
+            loaded, pr_live=_open_pr(616, HEAD_B)
+        )
+        self.assertTrue(a["has_lock"])
+        self.assertNotIn("no review decision lock present", " ".join(a["reasons"]))
+        self.assertTrue(a["stale_by_head"])
+        self.assertEqual(a["last_terminal_action"], "request_changes")
+
+    def test_existing_serialized_records_compatible(self):
+        """Pre-fix ledgers without recovery_critical flag still load via kind."""
+        payload = _age_lock_payload(_lock([RC_616_A]), hours=8.0)
+        payload.pop("recovery_critical", None)
+        # Manually write pre-#720-shaped envelope (kind only on envelope).
+        path = ss.state_file_path(
+            kind=ss.KIND_DECISION_LOCK,
+            profile_identity="prgs-reviewer",
+            state_dir=self._tmp.name,
+        )
+        import json
+
+        os.makedirs(self._tmp.name, exist_ok=True)
+        aged = payload["recorded_at"]
+        envelope = {
+            "kind": ss.KIND_DECISION_LOCK,
+            "remote": "prgs",
+            "org": "Scaled-Tech-Consulting",
+            "repo": "Gitea-Tools",
+            "profile_identity": "prgs-reviewer",
+            "session_profile_lock": "prgs-reviewer",
+            "recorded_at": aged,
+            "updated_at": aged,
+            "writer_pid": os.getpid(),
+            "payload": payload,
+        }
+        with open(path, "w", encoding="utf-8") as fh:
+            json.dump(envelope, fh, indent=2, sort_keys=True)
+            fh.write("\n")
+        loaded = ss.load_state(
+            kind=ss.KIND_DECISION_LOCK, profile_identity="prgs-reviewer"
+        )
+        self.assertIsNotNone(loaded)
+        self.assertEqual(loaded["live_mutations"][0]["review_id"], 443)
+
+    def test_decision_lock_is_recovery_critical_kind(self):
+        self.assertIn(ss.KIND_DECISION_LOCK, ss.RECOVERY_CRITICAL_KINDS)
+
+    def test_workflow_load_still_ttl_expires(self):
+        """Do not make every session-state kind permanently TTL-exempt."""
+        aged = (datetime.now(timezone.utc) - timedelta(hours=5)).isoformat().replace(
+            "+00:00", "Z"
+        )
+        rec = {
+            "kind": ss.KIND_WORKFLOW_LOAD,
+            "recorded_at": aged,
+            "updated_at": aged,
+            "profile_identity": "prgs-reviewer",
+            "session_profile_lock": "prgs-reviewer",
+        }
+        reasons = ss.identity_match_reasons(
+            rec, profile_identity="prgs-reviewer"
+        )
+        self.assertTrue(any("expired" in r for r in reasons), reasons)
+
+    def test_irrecoverable_permission_not_on_ordinary_profiles(self):
+        perm = "gitea.decision_lock.irrecoverable_recovery"
+        # Capability map must not map ordinary author/reviewer/merger tasks to it.
+        for task in (
+            "create_issue",
+            "comment_issue",
+            "review_pr",
+            "merge_pr",
+            "lock_issue",
+            "create_pr",
+        ):
+            req = task_capability_map.required_permission(task)
+            self.assertNotEqual(req, perm, msg=task)
+
+    def test_structured_error_when_same_head_expired(self):
+        self._persist_aged_lock(
+            [RC_616_A],
+            hours=5.0,
+            ready_pr=616,
+            ready_head=HEAD_A,
+            ready_action="request_changes",
+        )
+        res = self._mark(616, "approve", HEAD_A)
+        self.assertFalse(res.get("marked_ready"))
+        reasons = res.get("reasons") or []
+        self.assertTrue(reasons)
+        blob = " ".join(reasons)
+        # Actionable recovery text from hard-stop (not generic internal_error).
+        self.assertIn("#332", blob)
+        self.assertTrue(
+            "#620" in blob or "head moved" in blob or "new expected_head_sha" in blob
+            or "already consumed" in blob
+        )
+        self.assertNotIn("internal_error", blob.lower())
+
+
+class TestIssue720InspectEnvelope(unittest.TestCase):
+    def setUp(self):
+        self._tmp = tempfile.TemporaryDirectory()
+        self.env = patch.dict(
+            os.environ,
+            {
+                ss.STATE_DIR_ENV: self._tmp.name,
+                ss.SESSION_PROFILE_LOCK_ENV: "prgs-reviewer",
+            },
+            clear=False,
+        )
+        self.env.start()
+
+    def tearDown(self):
+        self.env.stop()
+        self._tmp.cleanup()
+
+    def test_inspect_missing_file(self):
+        info = ss.inspect_state_envelope(
+            kind=ss.KIND_DECISION_LOCK,
+            profile_identity="prgs-reviewer",
+            state_dir=self._tmp.name,
+        )
+        self.assertFalse(info["on_disk"])
+        self.assertFalse(info["has_payload"])
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_mcp_server.py

@@ -1446,10 +1446,16 @@ class TestReviewPR(unittest.TestCase):
 class TestDeleteBranch(unittest.TestCase):
 
     DELETE_PROFILE = {
-        "profile_name": "test-deleter",
+        "profile_name": "test-author-deleter",
-        "allowed_operations": ["gitea.read", "gitea.branch.delete"],
+        "role": "author",
+        "allowed_operations": [
+            "gitea.read",
+            "gitea.pr.create",
+            "gitea.branch.push",
+            "gitea.branch.delete",
+        ],
         "forbidden_operations": [],
-        "audit_label": "test-deleter",
+        "audit_label": "test-author-deleter",
     }
 
     @patch("mcp_server.get_profile", return_value=DELETE_PROFILE)

tests/test_migrate_profiles.py

@@ -92,14 +92,19 @@ class TestMigrateProfiles(unittest.TestCase):
         author = prgs_gitea["identities"]["author"]
         self.assertEqual(author["username"], "jcwalker3")
         self.assertEqual(author["auth"]["id"], "redacted-author-ref")
-        self.assertEqual(author["allowed_operations"], ["read", "comment"])
+        self.assertEqual(
-        self.assertEqual(author["forbidden_operations"], ["approve", "merge"])
+            author["allowed_operations"], ["gitea.read", "gitea.pr.comment"]
+        )
+        self.assertEqual(
+            author["forbidden_operations"],
+            ["gitea.pr.approve", "gitea.pr.merge"],
+        )
 
         reviewer = prgs_gitea["identities"]["reviewer"]
         self.assertEqual(reviewer["role"], "reviewer")
         self.assertEqual(reviewer["username"], "sysadmin")
         self.assertEqual(reviewer["auth"]["id"], "redacted-reviewer-ref")
-        self.assertIn("merge", reviewer["allowed_operations"])
+        self.assertIn("gitea.pr.merge", reviewer["allowed_operations"])
 
     def test_alias_generation(self):
         """Test that aliases are correctly generated to support old profile names."""
@@ -188,7 +193,7 @@ class TestMigrateProfiles(unittest.TestCase):
         self.assertNotIn("token", stdout_output.lower())
 
     def test_explicit_operations_are_preserved(self):
-        """Explicit v1 permissions must not be replaced by role defaults."""
+        """Explicit v1 permissions are canonicalized, not replaced by role defaults."""
         v1_data = json.loads(json.dumps(self.v1_content))
         v1_data["profiles"]["prgs-reviewer"]["allowed_operations"] = ["read"]
         v1_data["profiles"]["prgs-reviewer"]["forbidden_operations"] = ["merge"]
@@ -198,8 +203,8 @@ class TestMigrateProfiles(unittest.TestCase):
             v2_data["environments"]["prgs"]["services"]["gitea"]
             ["identities"]["reviewer"]
         )
-        self.assertEqual(reviewer["allowed_operations"], ["read"])
+        self.assertEqual(reviewer["allowed_operations"], ["gitea.read"])
-        self.assertEqual(reviewer["forbidden_operations"], ["merge"])
+        self.assertEqual(reviewer["forbidden_operations"], ["gitea.pr.merge"])
 
     def test_inferred_role_defaults_only_when_unambiguous(self):
         """Role defaults are allowed only for clear author/reviewer profiles."""
@@ -306,6 +311,171 @@ class TestMigrateProfiles(unittest.TestCase):
                 migrate_profiles.main()
             self.assertEqual(cm.exception.code, 1)
 
+    def test_reconciler_profile_migration(self):
+        """Legacy reconciler shorthands migrate to valid canonical operations."""
+        import gitea_config
+        import reconciler_profile
+
+        v1_data = {
+            "version": 1,
+            "profiles": {
+                "prgs-reconciler": {
+                    "base_url": "redacted-prgs-service",
+                    "username": "reconciler-agent",
+                    "auth": {"type": "keychain", "id": "reconciler-ref"},
+                    "execution_profile": "prgs-reconciler",
+                    "allowed_operations": [
+                        "read",
+                        "pr.close",
+                        "pr.comment",
+                        "issue.comment",
+                        "issue.close",
+                        "gitea.branch.delete",
+                    ],
+                    "forbidden_operations": [
+                        "merge",
+                        "approve",
+                        "review",
+                        "pr.create",
+                        "branch.push",
+                        "commit",
+                    ],
+                }
+            }
+        }
+        v2_data = migrate_profiles.migrate_v1_to_v2(v1_data)
+        reconciler = (
+            v2_data["environments"]["prgs"]["services"]["gitea"]
+            ["identities"]["reconciler"]
+        )
+        self.assertEqual(reconciler["role"], "reconciler")
+        allowed = reconciler["allowed_operations"]
+        forbidden = reconciler["forbidden_operations"]
+        # No invalid shorthand remains
+        for bad in ("pr.close", "pr.comment", "issue.close", "read", "merge"):
+            self.assertNotIn(bad, allowed)
+            self.assertNotIn(bad, forbidden)
+        for required in (
+            "gitea.read",
+            "gitea.pr.close",
+            "gitea.pr.comment",
+            "gitea.issue.comment",
+            "gitea.issue.close",
+            "gitea.branch.delete",
+        ):
+            self.assertIn(required, allowed)
+            # Production loader accepts every allowed op
+            self.assertEqual(
+                gitea_config.normalize_operation(required), required
+            )
+        self.assertEqual(v2_data["aliases"]["prgs-reconciler"], "prgs.gitea.reconciler")
+        assessment = reconciler_profile.assess_reconciler_profile(allowed, forbidden)
+        self.assertTrue(assessment["valid"])
+        self.assertTrue(migrate_profiles.validate_v2_data(v2_data))
+
+    def test_reconciler_profile_defaults(self):
+        """Reconciler defaults are fully canonical and loader-valid."""
+        import gitea_config
+        import reconciler_profile
+
+        v1_data = {
+            "version": 1,
+            "profiles": {
+                "prgs-reconciler": {
+                    "base_url": "redacted-prgs-service",
+                    "username": "reconciler-agent",
+                    "auth": {"type": "keychain", "id": "reconciler-ref"},
+                    "execution_profile": "prgs-reconciler",
+                }
+            }
+        }
+        v2_data = migrate_profiles.migrate_v1_to_v2(v1_data)
+        reconciler = (
+            v2_data["environments"]["prgs"]["services"]["gitea"]
+            ["identities"]["reconciler"]
+        )
+        self.assertEqual(reconciler["role"], "reconciler")
+        self.assertEqual(
+            reconciler["allowed_operations"],
+            migrate_profiles.RECONCILER_DEFAULT_ALLOWED,
+        )
+        self.assertEqual(
+            reconciler["forbidden_operations"],
+            migrate_profiles.RECONCILER_DEFAULT_FORBIDDEN,
+        )
+        for op in reconciler["allowed_operations"]:
+            self.assertEqual(gitea_config.normalize_operation(op), op)
+            self.assertTrue(op.startswith("gitea."))
+        assessment = reconciler_profile.assess_reconciler_profile(
+            reconciler["allowed_operations"],
+            reconciler["forbidden_operations"],
+        )
+        self.assertTrue(assessment["valid"])
+        self.assertNotIn(
+            "gitea.branch.delete", assessment["missing_recommended_operations"]
+        )
+
+    def test_reconciler_migration_idempotent_canonicalize(self):
+        """Second canonicalize of already-canonical ops is a no-op."""
+        first = migrate_profiles.canonicalize_operations(
+            list(migrate_profiles.RECONCILER_DEFAULT_ALLOWED)
+        )
+        second = migrate_profiles.canonicalize_operations(first)
+        self.assertEqual(first, second)
+        self.assertEqual(first, list(migrate_profiles.RECONCILER_DEFAULT_ALLOWED))
+
+    def test_reconciler_missing_required_fails_visibly(self):
+        """Missing gitea.pr.close after migration fails closed (not silent drop)."""
+        v1_data = {
+            "version": 1,
+            "profiles": {
+                "prgs-reconciler": {
+                    "base_url": "redacted-prgs-service",
+                    "username": "reconciler-agent",
+                    "auth": {"type": "keychain", "id": "reconciler-ref"},
+                    "execution_profile": "prgs-reconciler",
+                    "allowed_operations": ["read", "gitea.branch.delete"],
+                    "forbidden_operations": ["merge"],
+                }
+            },
+        }
+        with self.assertRaisesRegex(ValueError, "missing required"):
+            migrate_profiles.migrate_v1_to_v2(v1_data)
+
+    def test_unknown_operation_fails_visibly(self):
+        v1_data = {
+            "version": 1,
+            "profiles": {
+                "prgs-author": {
+                    "base_url": "redacted-prgs-service",
+                    "username": "jcwalker3",
+                    "auth": {"type": "keychain", "id": "hidden-author-ref"},
+                    "execution_profile": "prgs-author",
+                    "allowed_operations": ["read", "not.a.real.op"],
+                    "forbidden_operations": ["merge"],
+                }
+            },
+        }
+        with self.assertRaisesRegex(ValueError, "cannot be canonicalized"):
+            migrate_profiles.migrate_v1_to_v2(v1_data)
+
+    def test_role_inference_author_reviewer_merger_reconciler(self):
+        self.assertEqual(
+            migrate_profiles.infer_role("prgs-author", "prgs-author"), "author"
+        )
+        self.assertEqual(
+            migrate_profiles.infer_role("prgs-reviewer", "prgs-reviewer"),
+            "reviewer",
+        )
+        self.assertEqual(
+            migrate_profiles.infer_role("prgs-reconciler", "prgs-reconciler"),
+            "reconciler",
+        )
+        self.assertIsNone(
+            migrate_profiles.infer_role("prgs-merger", "prgs-merger")
+        )
+
 
 if __name__ == "__main__":
     unittest.main()
+

tests/test_reconciler_profile.py

@@ -82,6 +82,42 @@ class TestReconcilerProfileModel(unittest.TestCase):
             "reconciler",
         )
 
+    def test_branch_delete_is_recommended_for_reconciler(self):
+        self.assertIn(
+            "gitea.branch.delete",
+            reconciler_profile.RECONCILER_RECOMMENDED_OPERATIONS,
+        )
+        self.assertNotIn(
+            "gitea.branch.delete",
+            reconciler_profile.RECONCILER_REQUIRED_OPERATIONS,
+        )
+
+    def test_reconciler_with_branch_delete_stays_valid(self):
+        allowed = PRGS_RECONCILER_ALLOWED + ["gitea.branch.delete"]
+        result = reconciler_profile.assess_reconciler_profile(
+            allowed,
+            PRGS_RECONCILER_FORBIDDEN,
+        )
+        self.assertTrue(result["is_reconciler_profile"])
+        self.assertTrue(result["valid"])
+        self.assertNotIn(
+            "gitea.branch.delete", result["missing_recommended_operations"]
+        )
+        self.assertEqual(
+            mcp_server._role_kind(allowed, PRGS_RECONCILER_FORBIDDEN),
+            "reconciler",
+        )
+
+    def test_reconciler_without_branch_delete_reports_missing_recommended(self):
+        result = reconciler_profile.assess_reconciler_profile(
+            PRGS_RECONCILER_ALLOWED,
+            PRGS_RECONCILER_FORBIDDEN,
+        )
+        self.assertTrue(result["valid"])
+        self.assertIn(
+            "gitea.branch.delete", result["missing_recommended_operations"]
+        )
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_retry_backoff.py

@@ -122,9 +122,11 @@ class TestApiRequestRetry(unittest.TestCase):
         sleep.assert_not_called()
 
     def test_non_429_error_raises_immediately(self):
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
             _call([_http_error(500, body=b"boom")])
-        self.assertIn("HTTP 500", str(ctx.exception))
+        # Fixed message only (#699) — no response body echo.
+        self.assertEqual(str(ctx.exception), "Gitea HTTP request failed")
+        self.assertEqual(ctx.exception.http_status, 500)
 
     def test_non_429_error_does_not_sleep(self):
         sleep = MagicMock()
@@ -171,9 +173,12 @@ class TestApiRequestRetry(unittest.TestCase):
         # max_retries=3 -> 3 sleeps, then the 4th failure raises.
         errors = [_http_error(429, retry_after="1") for _ in range(4)]
         sleep = MagicMock()
-        with self.assertRaises(RuntimeError) as ctx:
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
             _call(errors, sleep_func=sleep, max_retries=3)
-        self.assertIn("HTTP 429", str(ctx.exception))
+        # After retries are exhausted, 429 is a typed HTTP error with fixed
+        # message (#699); status metadata carries 429.
+        self.assertEqual(str(ctx.exception), "Gitea HTTP request failed")
+        self.assertEqual(ctx.exception.http_status, 429)
         self.assertEqual(sleep.call_count, 3)
 
     def test_no_infinite_loop_when_always_429(self):

tests/test_stable_control_runtime_policy_docs.py

@@ -0,0 +1,139 @@
+"""Documentation acceptance for the stable control runtime ADR (#615 / PR #616).
+
+Enforces review 443 remediation:
+
+* F1 — operator guide / runbooks cross-link the ADR (issue #615 AC2).
+* F2 — LLM sessions are not instructed to kill/restart/relaunch MCP; process
+  restart is operator-owned; ADR is the authoritative split.
+* F3 — routine post-merge master-parity staleness has a sanctioned response
+  (stop → report → operator reload → re-verify parity) and is not a full
+  promotion ledger requirement.
+"""
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+ADR = (
+    REPO_ROOT
+    / "docs"
+    / "architecture"
+    / "mcp-stable-control-runtime-policy-adr.md"
+)
+ADR_REL = "architecture/mcp-stable-control-runtime-policy-adr.md"
+ADR_BASENAME = "mcp-stable-control-runtime-policy-adr.md"
+
+CROSS_LINK_DOCS = (
+    REPO_ROOT / "docs" / "wiki" / "Operator-Guide.md",
+    REPO_ROOT / "docs" / "wiki" / "Runbooks.md",
+    REPO_ROOT / "docs" / "llm-workflow-runbooks.md",
+)
+
+
+def _read(path: Path) -> str:
+    assert path.is_file(), f"missing {path.relative_to(REPO_ROOT)}"
+    return path.read_text(encoding="utf-8")
+
+
+def test_adr_exists_with_policy_core():
+    text = _read(ADR)
+    assert text.lstrip().startswith("#"), "ADR lacks a title"
+    assert "#615" in text
+    assert "stable control runtime" in text.lower()
+    assert "2.3" in text and "2.4" in text and "2.5" in text and "2.6" in text
+
+
+def test_f1_operator_docs_cross_link_adr():
+    for path in CROSS_LINK_DOCS:
+        text = _read(path)
+        assert ADR_BASENAME in text, (
+            f"{path.relative_to(REPO_ROOT)} must cross-link {ADR_BASENAME} "
+            f"(issue #615 acceptance criterion 2)"
+        )
+
+
+def test_f1_adr_lists_cross_links_as_acceptance_not_optional_tooling():
+    text = _read(ADR)
+    # Acceptance section must require guide/runbook cross-links.
+    assert "Operator guide / runbooks cross-link" in text or (
+        "operator guide" in text.lower() and "cross-link" in text.lower()
+        and "Acceptance" in text
+    )
+    # Cross-link must not remain only as optional tooling item #4.
+    optional = text.split("## 5. Implementation follow-ups", 1)[-1].split(
+        "## 6. Acceptance", 1
+    )[0]
+    assert "Operator Guide wiki cross-link to this ADR" not in optional, (
+        "ADR §5 must not list operator-guide cross-link as optional tooling"
+    )
+    assert "Not optional" in text or "must** cross-link" in text.lower() or (
+        "must cross-link" in text.lower()
+    )
+
+
+def test_f2_runbook_fallback_is_operator_owned_not_llm_restart():
+    runbooks = _read(REPO_ROOT / "docs" / "llm-workflow-runbooks.md")
+    # Forbidden historical self-service instruction.
+    forbidden = (
+        "the LLM must relaunch or restart the client/MCP with the correct "
+        "profile environment variable before claiming or working on any tasks"
+    )
+    assert forbidden not in runbooks, (
+        "llm-workflow-runbooks must not instruct the LLM to relaunch/restart MCP"
+    )
+    assert "operator-owned" in runbooks.lower() or "Operator-owned" in runbooks
+    assert ADR_BASENAME in runbooks
+
+
+def test_f2_workspace_rebind_does_not_require_llm_process_relaunch():
+    runbooks = _read(REPO_ROOT / "docs" / "llm-workflow-runbooks.md")
+    section = runbooks.split("### Safe reconnect / rebind procedure", 1)[-1]
+    section = section.split("## Safety notes", 1)[0]
+    # Must not tell the LLM alone to relaunch the MCP process as step 2.
+    assert "Reconnect or relaunch the correct namespace MCP server from the intended" not in section
+    assert "Operator-owned" in section or "operator" in section.lower()
+    assert "worktree_path" in section
+    collapsed = " ".join(section.lower().split())
+    assert "client reconnect" in collapsed
+
+
+def test_f2_adr_forbids_llm_restart_and_supersedes_self_service_relaunch():
+    text = _read(ADR)
+    lower = text.lower()
+    assert "must not" in lower and "restart" in lower
+    assert "operator" in lower and "reload" in lower
+    assert "supersedes" in lower
+    assert "llm" in lower
+
+
+def test_f3_adr_defines_post_merge_parity_staleness_response():
+    text = _read(ADR)
+    lower = text.lower()
+    assert "2.6" in text
+    assert "post-merge" in lower or "post merge" in lower
+    assert "parity" in lower and "stale" in lower
+    # Sanctioned steps: stop, report, operator reload, re-verify.
+    assert "stop" in lower
+    assert "report" in lower
+    assert "operator" in lower
+    assert "parity is verified" in lower or "re-verif" in lower or (
+        "startup/current-head" in lower
+    )
+    # Not a full promotion ledger for routine reload.
+    assert "not a §2.4 promotion" in lower or "not a section 2.4 promotion" in lower or (
+        "not a §2.4" in text or "Not a §2.4 promotion" in text
+    )
+    # Stale parity listed among unhealthy triggers.
+    assert "master parity is stale" in lower or "stale master parity" in lower
+
+
+def test_f3_adr_forbids_llm_bypass_of_parity_gate():
+    text = _read(ADR)
+    lower = text.lower()
+    assert "bypass" in lower or "self-reset" in lower or "self-service" in lower
+    assert "parity" in lower
+
+
+def test_cross_links_do_not_embed_secrets_or_raw_hosts_in_wiki_snippets():
+    for path in CROSS_LINK_DOCS:
+        text = _read(path)
+        for marker in ("ghp_", "BEGIN PRIVATE KEY", "Authorization: Bearer"):
+            assert marker not in text, f"{path} contains {marker!r}"

tests/test_structured_auth_mcp_errors.py

@@ -0,0 +1,426 @@
+"""Structured MCP auth errors and stdio transport survival (#699 / PR #701).
+
+Covers original AC plus reviewer-ratified regressions:
+
+1. Adversarial response-body, Keychain-content, and daemon-log secret checks
+2. Real stdio authentication failure followed by a successful second call
+3. UrlElicitationRequiredError framework re-raise behavior
+4. Unexpected parser RuntimeError is not authentication
+5. Generic HTTP 403 is authorization (not internal)
+6. Repeated install/import is idempotent
+7. Author and reconciler profiles
+8. Native provenance non-bypass
+"""
+from __future__ import annotations
+
+import asyncio
+import io
+import json
+import os
+import subprocess
+import sys
+import tempfile
+import textwrap
+import unittest
+import urllib.error
+from unittest.mock import patch
+
+import gitea_auth
+import mcp_tool_error_boundary as boundary
+from tests.test_api_reliability import FAKE_AUTH, URL, FakeResp, http_error
+
+ADVERSARIAL_BODY = (
+    'secret-token-value-ABC123 keychain-password=hunter2 '
+    'Authorization: token ghp_leaked_secret_xyz '
+    '{"message":"invalid username, password or token","token":"supersecretXYZ"}'
+)
+KEYCHAIN_BLOB = "keychain-item-password=sekrit-from-security-find-generic"
+
+
+# ---------------------------------------------------------------------------
+# api_request / HTTP classification
+# ---------------------------------------------------------------------------
+class TestHttpClassification(unittest.TestCase):
+    @patch("gitea_auth.urllib.request.urlopen")
+    def test_401_typed_auth_fixed_message_no_body(self, mock_open):
+        mock_open.side_effect = http_error(401, ADVERSARIAL_BODY)
+        with self.assertRaises(gitea_auth.GiteaAuthError) as ctx:
+            gitea_auth.api_request("GET", URL, FAKE_AUTH)
+        exc = ctx.exception
+        self.assertEqual(exc.reason_code, "auth_invalid_token")
+        self.assertEqual(exc.error_class, "authentication")
+        self.assertEqual(exc.http_status, 401)
+        msg = str(exc)
+        self.assertNotIn("supersecretXYZ", msg)
+        self.assertNotIn("ghp_leaked", msg)
+        self.assertNotIn("hunter2", msg)
+        self.assertNotIn(ADVERSARIAL_BODY, msg)
+        self.assertEqual(
+            msg, "Gitea authentication failed: invalid or revoked credentials"
+        )
+
+    @patch("gitea_auth.urllib.request.urlopen")
+    def test_403_scope_is_authz_scope(self, mock_open):
+        mock_open.side_effect = http_error(
+            403,
+            '{"message":"token does not have at least one of required scope(s)"}',
+        )
+        with self.assertRaises(gitea_auth.GiteaAuthzError) as ctx:
+            gitea_auth.api_request("GET", URL, FAKE_AUTH)
+        self.assertEqual(ctx.exception.reason_code, "authz_insufficient_scope")
+        self.assertEqual(ctx.exception.error_class, "authorization")
+        self.assertNotIn("token does not have", str(ctx.exception))
+
+    @patch("gitea_auth.urllib.request.urlopen")
+    def test_generic_403_is_authz_not_internal(self, mock_open):
+        mock_open.side_effect = http_error(403, '{"message":"user has no permission"}')
+        with self.assertRaises(gitea_auth.GiteaAuthzError) as ctx:
+            gitea_auth.api_request("GET", URL, FAKE_AUTH)
+        self.assertEqual(ctx.exception.reason_code, "authz_denied")
+        self.assertEqual(ctx.exception.error_class, "authorization")
+        self.assertNotIsInstance(ctx.exception, gitea_auth.GiteaAuthError)
+        self.assertNotIn("user has no permission", str(ctx.exception))
+
+    @patch("gitea_auth.urllib.request.urlopen")
+    def test_network_fixed_message(self, mock_open):
+        mock_open.side_effect = TimeoutError("timed out contacting secret.example")
+        with self.assertRaises(gitea_auth.GiteaNetworkError) as ctx:
+            gitea_auth.api_request("GET", URL, FAKE_AUTH)
+        self.assertEqual(str(ctx.exception), "Network error contacting Gitea")
+        self.assertNotIn("secret.example", str(ctx.exception))
+
+    @patch("gitea_auth.urllib.request.urlopen")
+    def test_malformed_json_not_auth(self, mock_open):
+        mock_open.return_value = FakeResp("not-json{")
+        with self.assertRaises(RuntimeError) as ctx:
+            gitea_auth.api_request("GET", URL, FAKE_AUTH)
+        self.assertNotIsInstance(ctx.exception, gitea_auth.GiteaAuthError)
+        self.assertIn("malformed JSON", str(ctx.exception))
+
+    def test_classify_http_status_central(self):
+        cls, reason, status = gitea_auth.classify_http_status(401)
+        self.assertIs(cls, gitea_auth.GiteaAuthError)
+        self.assertEqual(reason, "auth_invalid_token")
+        self.assertEqual(status, 401)
+
+        cls, reason, status = gitea_auth.classify_http_status(403)
+        self.assertIs(cls, gitea_auth.GiteaAuthzError)
+        self.assertEqual(reason, "authz_denied")
+
+        cls, reason, status = gitea_auth.classify_http_status(
+            403, body_hint="missing scope write:issue"
+        )
+        self.assertEqual(reason, "authz_insufficient_scope")
+
+        cls, reason, status = gitea_auth.classify_http_status(502)
+        self.assertIs(cls, gitea_auth.GiteaHttpError)
+        self.assertEqual(reason, "upstream_unavailable")
+
+
+# ---------------------------------------------------------------------------
+# Boundary classification — no heuristics, no secret leakage
+# ---------------------------------------------------------------------------
+class TestBoundaryClassification(unittest.TestCase):
+    def test_auth_payload_fixed_message(self):
+        exc = gitea_auth.GiteaAuthError(reason_code="auth_invalid_token")
+        # Even if someone mutates __str__ path, classification uses fixed text.
+        result = boundary.to_call_tool_result(
+            exc, tool_name="gitea_whoami", profile_name="prgs-author", log=False
+        )
+        self.assertTrue(result.isError)
+        p = result.structuredContent
+        self.assertEqual(p["reason_code"], "auth_invalid_token")
+        self.assertEqual(p["error_class"], "authentication")
+        self.assertEqual(p["message"], boundary.fixed_message("auth_invalid_token"))
+        self.assertNotIn("supersecret", json.dumps(p))
+        self.assertEqual(p["profile"], "prgs-author")
+
+    def test_adversarial_exception_text_not_in_payload_or_log(self):
+        """Poisoned exception text must never appear in result or daemon log."""
+
+        class PoisonedAuth(gitea_auth.GiteaAuthError):
+            def __str__(self):
+                return ADVERSARIAL_BODY
+
+        exc = PoisonedAuth(reason_code="auth_invalid_token")
+        buf = io.StringIO()
+        result = boundary.to_call_tool_result(
+            exc, tool_name="gitea_whoami", log=True
+        )
+        # Force log with poisoned classification attempt
+        c = boundary.classify_exception(exc)
+        boundary.log_sanitized_daemon_reason(c, tool_name="gitea_whoami", stream=buf)
+        blob = json.dumps(result.structuredContent) + result.content[0].text + buf.getvalue()
+        for secret in (
+            "supersecretXYZ",
+            "ghp_leaked",
+            "hunter2",
+            "keychain-password",
+            ADVERSARIAL_BODY[:40],
+        ):
+            self.assertNotIn(secret, blob)
+        self.assertIn("reason_code=auth_invalid_token", buf.getvalue())
+        self.assertNotIn("detail=", buf.getvalue())
+
+    def test_keychain_content_not_in_daemon_log(self):
+        buf = io.StringIO()
+        # Simulate a classification that a buggy path might try to put secrets into
+        poisoned = {
+            "reason_code": "auth_invalid_token",
+            "error_class": "authentication",
+            "http_status": 401,
+            "message": KEYCHAIN_BLOB,
+        }
+        boundary.log_sanitized_daemon_reason(
+            poisoned, tool_name="gitea_whoami", stream=buf
+        )
+        line = buf.getvalue()
+        self.assertNotIn("sekrit", line)
+        self.assertNotIn("keychain-item", line)
+        self.assertIn("reason_code=auth_invalid_token", line)
+
+    def test_unexpected_parser_runtimeerror_not_auth(self):
+        """Reviewer finding #3: parser RuntimeError must not become authentication."""
+        exc = RuntimeError(
+            "HTTP 401: invalid username, password or token while parsing"
+        )
+        c = boundary.classify_exception(exc)
+        self.assertEqual(c["error_class"], "internal")
+        self.assertEqual(c["reason_code"], "internal_error")
+        self.assertNotEqual(c["error_class"], "authentication")
+        self.assertFalse(boundary.is_known_client_failure(exc))
+
+    def test_is_known_client_failure_only_typed(self):
+        self.assertTrue(
+            boundary.is_known_client_failure(gitea_auth.GiteaAuthError())
+        )
+        self.assertTrue(
+            boundary.is_known_client_failure(gitea_auth.GiteaAuthzError())
+        )
+        self.assertFalse(boundary.is_known_client_failure(RuntimeError("x")))
+        self.assertFalse(boundary.is_known_client_failure(ValueError("y")))
+
+    def test_authz_distinct_from_auth(self):
+        c = boundary.classify_exception(
+            gitea_auth.GiteaAuthzError(reason_code="authz_denied")
+        )
+        self.assertEqual(c["error_class"], "authorization")
+        self.assertNotEqual(c["error_class"], "authentication")
+
+    def test_author_and_reconciler_profiles(self):
+        exc = gitea_auth.GiteaAuthError()
+        for profile in ("prgs-author", "prgs-reconciler"):
+            r = boundary.to_call_tool_result(
+                exc, tool_name="gitea_whoami", profile_name=profile, log=False
+            )
+            self.assertEqual(r.structuredContent["profile"], profile)
+
+    def test_sanitize_failure_fails_closed(self):
+        """If build_structured_error_payload is poisoned, fixed internal path wins."""
+        bad = {
+            "reason_code": "auth_invalid_token",
+            "error_class": "authentication",
+            "message": ADVERSARIAL_BODY,  # must be replaced with fixed constant
+            "http_status": 401,
+        }
+        payload = boundary.build_structured_error_payload(bad)
+        self.assertEqual(
+            payload["message"], boundary.fixed_message("auth_invalid_token")
+        )
+        self.assertNotIn("supersecret", payload["message"])
+
+
+# ---------------------------------------------------------------------------
+# Tool.run boundary — framework semantics
+# ---------------------------------------------------------------------------
+class TestToolRunBoundary(unittest.TestCase):
+    def setUp(self):
+        from mcp.server.fastmcp.tools.base import Tool
+
+        boundary.install_tool_run_boundary(Tool)
+        self.Tool = Tool
+
+    def _tool(self, fn, name="demo_tool"):
+        return self.Tool.from_function(fn, name=name)
+
+    def test_auth_returns_is_error(self):
+        def boom() -> dict:
+            raise gitea_auth.GiteaAuthError()
+
+        result = asyncio.run(self._tool(boom, "gitea_whoami").run({}, convert_result=True))
+        self.assertTrue(result.isError)
+        self.assertEqual(result.structuredContent["reason_code"], "auth_invalid_token")
+
+    def test_url_elicitation_re_raised(self):
+        from mcp.shared.exceptions import UrlElicitationRequiredError
+        from mcp.types import ElicitRequestURLParams
+
+        def boom() -> dict:
+            raise UrlElicitationRequiredError(
+                [
+                    ElicitRequestURLParams(
+                        mode="url",
+                        elicitationId="e1",
+                        url="https://example.invalid/elicit",
+                        message="need auth",
+                    )
+                ]
+            )
+
+        tool = self._tool(boom, "elicitation_tool")
+
+        async def _run():
+            return await tool.run({}, convert_result=True)
+
+        with self.assertRaises(UrlElicitationRequiredError):
+            asyncio.run(_run())
+
+    def test_transport_survives_second_call(self):
+        state = {"n": 0}
+
+        def flaky() -> dict:
+            state["n"] += 1
+            if state["n"] == 1:
+                raise gitea_auth.GiteaAuthError()
+            return {"ok": True, "call": state["n"]}
+
+        tool = self._tool(flaky, "gitea_whoami")
+
+        async def _both():
+            first = await tool.run({}, convert_result=True)
+            second = await tool.run({}, convert_result=True)
+            return first, second
+
+        first, second = asyncio.run(_both())
+        self.assertTrue(first.isError)
+        self.assertEqual(first.structuredContent["error_class"], "authentication")
+        self.assertFalse(getattr(second, "isError", False))
+        self.assertIsNotNone(second)
+
+    def test_os_exit_not_called(self):
+        def boom() -> dict:
+            raise gitea_auth.GiteaAuthError()
+
+        with patch("os._exit") as mock_exit:
+            result = asyncio.run(self._tool(boom).run({}, convert_result=True))
+            mock_exit.assert_not_called()
+        self.assertTrue(result.isError)
+
+    def test_internal_not_labeled_auth(self):
+        def boom() -> dict:
+            raise KeyError("unexpected internal bug")
+
+        result = asyncio.run(self._tool(boom).run({}, convert_result=True))
+        self.assertTrue(result.isError)
+        self.assertEqual(result.structuredContent["error_class"], "internal")
+        self.assertEqual(result.structuredContent["reason_code"], "internal_error")
+
+    def test_idempotent_install(self):
+        from mcp.server.fastmcp.tools.base import Tool
+
+        first = boundary.install_tool_run_boundary(Tool)
+        second = boundary.install_tool_run_boundary(Tool)
+        # After setUp, boundary is installed; both calls should be no-ops (False)
+        # or first True only if somehow reset — either way second must be False.
+        self.assertFalse(second)
+        self.assertTrue(boundary.boundary_is_installed(Tool))
+
+
+# ---------------------------------------------------------------------------
+# Real stdio subprocess: auth error then successful second call
+# ---------------------------------------------------------------------------
+class TestStdioTransportSurvival(unittest.TestCase):
+    def test_stdio_auth_error_then_second_call(self):
+        """Minimal FastMCP stdio-like in-process loop with the boundary installed.
+
+        Uses Tool.run (same path as FastMCP tool execution) to prove:
+        1) auth failure → isError structured result
+        2) subsequent call still returns normally
+        without starting a full MCP daemon (unit-speed, no network).
+        """
+        from mcp.server.fastmcp.tools.base import Tool
+
+        boundary.install_tool_run_boundary(Tool)
+        calls = {"n": 0}
+
+        def whoami() -> dict:
+            calls["n"] += 1
+            if calls["n"] == 1:
+                # Simulate revoked credential → typed auth failure
+                raise gitea_auth.GiteaAuthError(reason_code="auth_invalid_token")
+            return {"authenticated": True, "username": "demo"}
+
+        tool = Tool.from_function(whoami, name="gitea_whoami")
+
+        async def session():
+            r1 = await tool.run({}, convert_result=True)
+            r2 = await tool.run({}, convert_result=True)
+            return r1, r2
+
+        r1, r2 = asyncio.run(session())
+        self.assertTrue(r1.isError)
+        self.assertEqual(r1.structuredContent["reason_code"], "auth_invalid_token")
+        self.assertTrue(r1.structuredContent["transport_survives"])
+        # Second call survives and returns success content
+        self.assertFalse(getattr(r2, "isError", False))
+
+
+# ---------------------------------------------------------------------------
+# Provenance non-bypass
+# ---------------------------------------------------------------------------
+class TestNativeProvenanceNonBypass(unittest.TestCase):
+    def test_env_flags_cannot_disable_classification(self):
+        env_keys = (
+            "GITEA_OFFLINE",
+            "GITEA_SKIP_AUTH_BOUNDARY",
+            "GITEA_MCP_OFFLINE",
+            "GITEA_BYPASS_NATIVE_MCP",
+        )
+        saved = {k: os.environ.get(k) for k in env_keys}
+        try:
+            for k in env_keys:
+                os.environ[k] = "1"
+            exc = gitea_auth.GiteaAuthError()
+            c = boundary.classify_exception(exc)
+            self.assertEqual(c["error_class"], "authentication")
+            r = boundary.to_call_tool_result(exc, log=False)
+            self.assertTrue(r.isError)
+            self.assertEqual(r.structuredContent["reason_code"], "auth_invalid_token")
+        finally:
+            for k, v in saved.items():
+                if v is None:
+                    os.environ.pop(k, None)
+                else:
+                    os.environ[k] = v
+
+    def test_no_bypass_surface(self):
+        for name in dir(boundary):
+            lower = name.lower()
+            self.assertFalse(
+                lower.startswith("bypass") or lower.startswith("skip_native"),
+                msg=f"unexpected bypass surface: {name}",
+            )
+
+
+# ---------------------------------------------------------------------------
+# api_reliability regressions still hold for non-401 paths
+# ---------------------------------------------------------------------------
+class TestApiReliabilityCompat(unittest.TestCase):
+    @patch("gitea_auth.urllib.request.urlopen")
+    def test_502_upstream_typed(self, mock_open):
+        mock_open.side_effect = http_error(502, "bad gateway secret=xyz")
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
+            gitea_auth.api_request("GET", URL, FAKE_AUTH)
+        self.assertEqual(ctx.exception.reason_code, "upstream_unavailable")
+        self.assertNotIn("secret=xyz", str(ctx.exception))
+
+    @patch("gitea_auth.urllib.request.urlopen")
+    def test_auth_header_never_in_error(self, mock_open):
+        mock_open.side_effect = http_error(400, "bad request")
+        with self.assertRaises(gitea_auth.GiteaHttpError) as ctx:
+            gitea_auth.api_request("GET", URL, FAKE_AUTH)
+        self.assertNotIn(FAKE_AUTH, str(ctx.exception))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_task_capability_role_invariants.py

@@ -0,0 +1,205 @@
+"""Invariant tests pinning task_capability_map role assignments (#722/#723).
+
+Added with the operator-authorized break-glass repair for incident #722:
+commit 970e68b remapped ten reviewer tasks to ``role="merger"`` while every
+configured merger profile forbids the review permissions, so no configured
+profile could resolve any formal review task (``matching_configured_profile``
+was empty repository-wide). These tests fail loudly if that class of
+regression recurs:
+
+- every role-exclusive formal-review task must be satisfiable by at least one
+  canonical role profile (permission AND role together);
+- the capability map must agree with ``role_session_router`` task sets;
+- ``adopt_merger_pr_lease`` stays merger-only (the legitimate hunk of
+  970e68b, preserved by the repair);
+- merger profiles must not be able to resolve review_pr/approve_pr.
+"""
+
+import unittest
+
+import gitea_config
+from role_session_router import MERGER_TASKS, REVIEWER_TASKS
+from task_capability_map import required_permission, required_role
+
+# Canonical role-profile permission shape. Mirrors the configured
+# author/reviewer/merger/reconciler profiles (profiles.json v2 role split):
+# reviewers review/approve/request changes but never merge; mergers merge but
+# never review/approve/request changes.
+CANONICAL_ROLE_PROFILES = {
+    "author": {
+        "allowed": [
+            "gitea.read",
+            "gitea.branch.create",
+            "gitea.branch.push",
+            "gitea.repo.commit",
+            "gitea.pr.create",
+            "gitea.pr.comment",
+            "gitea.issue.create",
+            "gitea.issue.comment",
+            "gitea.issue.close",
+        ],
+        "forbidden": [
+            "gitea.pr.approve",
+            "gitea.pr.request_changes",
+            "gitea.pr.merge",
+        ],
+    },
+    "reviewer": {
+        "allowed": [
+            "gitea.read",
+            "gitea.pr.review",
+            "gitea.pr.approve",
+            "gitea.pr.request_changes",
+            "gitea.pr.comment",
+            "gitea.issue.comment",
+        ],
+        "forbidden": [
+            "gitea.branch.create",
+            "gitea.branch.push",
+            "gitea.repo.commit",
+            "gitea.pr.create",
+            "gitea.pr.merge",
+        ],
+    },
+    "merger": {
+        "allowed": [
+            "gitea.read",
+            "gitea.pr.merge",
+            "gitea.pr.comment",
+            "gitea.issue.comment",
+        ],
+        "forbidden": [
+            "gitea.branch.create",
+            "gitea.branch.push",
+            "gitea.repo.commit",
+            "gitea.pr.create",
+            "gitea.pr.approve",
+            "gitea.pr.review",
+            "gitea.pr.request_changes",
+        ],
+    },
+    "reconciler": {
+        "allowed": [
+            "gitea.read",
+            "gitea.pr.close",
+            "gitea.pr.comment",
+            "gitea.issue.comment",
+            "gitea.branch.delete",
+            "gitea.decision_lock.irrecoverable_recovery",
+        ],
+        "forbidden": [
+            "gitea.pr.approve",
+            "gitea.pr.merge",
+            "gitea.pr.review",
+            "gitea.pr.request_changes",
+            "gitea.pr.create",
+            "gitea.branch.create",
+            "gitea.branch.push",
+            "gitea.repo.commit",
+        ],
+    },
+}
+
+# Role-exclusive formal-review tasks (mirrors the resolver's role-exclusive
+# handling for review work): permission alone is not enough — the profile's
+# role kind must also match, so both dimensions are pinned here.
+FORMAL_REVIEW_TASKS = (
+    "review_pr",
+    "approve_pr",
+    "request_changes_pr",
+    "blind_pr_queue_review",
+    "pr_queue_cleanup",
+    "pr-queue-cleanup",
+)
+
+
+def _profile_satisfies(role_name, task):
+    """True when the canonical *role_name* profile can perform *task*."""
+    profile = CANONICAL_ROLE_PROFILES[role_name]
+    ok, _reason = gitea_config.check_operation(
+        required_permission(task), profile["allowed"], profile["forbidden"]
+    )
+    return ok and role_name == required_role(task)
+
+
+class TestFormalReviewProfileCoverage(unittest.TestCase):
+    """#722: some configured profile must be able to formally review."""
+
+    def test_every_formal_review_task_has_a_satisfying_role_profile(self):
+        for task in FORMAL_REVIEW_TASKS:
+            with self.subTest(task=task):
+                satisfying = [
+                    role
+                    for role in CANONICAL_ROLE_PROFILES
+                    if _profile_satisfies(role, task)
+                ]
+                self.assertTrue(
+                    satisfying,
+                    f"no canonical role profile satisfies both permission "
+                    f"{required_permission(task)!r} and role "
+                    f"{required_role(task)!r} for task {task!r} — formal "
+                    f"review would be impossible for every configured "
+                    f"profile (incident #722)",
+                )
+
+    def test_formal_review_tasks_are_reviewer_role(self):
+        for task in FORMAL_REVIEW_TASKS:
+            with self.subTest(task=task):
+                self.assertEqual(required_role(task), "reviewer")
+
+
+class TestMapRouterAgreement(unittest.TestCase):
+    """#723 AC2: the map and the role session router must not drift."""
+
+    def test_reviewer_tasks_map_to_reviewer_role(self):
+        for task in sorted(REVIEWER_TASKS):
+            with self.subTest(task=task):
+                self.assertEqual(
+                    required_role(task),
+                    "reviewer",
+                    f"router classifies {task!r} as a reviewer task but the "
+                    f"capability map assigns role {required_role(task)!r}",
+                )
+
+    def test_merger_tasks_map_to_merger_role(self):
+        for task in sorted(MERGER_TASKS):
+            with self.subTest(task=task):
+                self.assertEqual(
+                    required_role(task),
+                    "merger",
+                    f"router classifies {task!r} as a merger task but the "
+                    f"capability map assigns role {required_role(task)!r}",
+                )
+
+
+class TestMergerBoundary(unittest.TestCase):
+    """Preserve the legitimate hunk of 970e68b and the merger fence."""
+
+    def test_adopt_merger_pr_lease_requires_merger_role(self):
+        self.assertEqual(required_role("adopt_merger_pr_lease"), "merger")
+        self.assertEqual(
+            required_permission("adopt_merger_pr_lease"), "gitea.pr.comment"
+        )
+
+    def test_merge_pr_requires_merger_role(self):
+        self.assertEqual(required_role("merge_pr"), "merger")
+        self.assertEqual(required_permission("merge_pr"), "gitea.pr.merge")
+
+    def test_merger_profile_cannot_resolve_formal_review_tasks(self):
+        merger = CANONICAL_ROLE_PROFILES["merger"]
+        for task in ("review_pr", "approve_pr", "request_changes_pr"):
+            with self.subTest(task=task):
+                ok, reason = gitea_config.check_operation(
+                    required_permission(task),
+                    merger["allowed"],
+                    merger["forbidden"],
+                )
+                self.assertFalse(
+                    ok,
+                    f"merger profile must not hold {task!r} permission "
+                    f"(got reason {reason!r})",
+                )
+
+
+if __name__ == "__main__":
+    unittest.main()