Merge commit '578c44b685a7ff5b01006c5e398bfac9863e0d8d' into feat/issue-661-drain-proof-hard-gate

The acks_or_timeout check treated an absent `acks` key as proof that no
session needed to acknowledge: `drain_state.get("acks") or {}` collapsed
absent, None, and empty into the same value, and the resulting empty mapping
satisfied `no_sessions_to_ack`. The impact report's counts.sessions_live_other
was never consulted, so absence of evidence was read as evidence of absence.

Reproduced at head 1cbbde0089: with
sessions_live_other = 3 and the acknowledgement key absent, acks_or_timeout
passed with detail "no other live sessions required to acknowledge", the proof
minted clean, and gate_apply_restart returned verdict allow — a restart
authorized against three live sessions with zero acknowledgement evidence, and
the resulting artifact carried a valid signature.

Whether acknowledgement is required is now derived from the impact report,
never from the shape of the drain state:

- _live_session_count() reads counts.sessions_live_other and returns None for a
  missing, malformed, negative, or bool value, so an unreadable report fails
  closed instead of reading as "nobody was live".
- Absent, None, non-mapping, empty, partially-covering, and unparseable or
  stale acknowledgement data all fail closed while live sessions require
  acknowledgement.
- _is_acknowledged() no longer coerces with str(); only an explicit
  "ack"/"acked"/"acknowledged" string counts, so None, timestamps, and
  "pending"/"stale" markers are never read as an acknowledgement.
- Present-but-unacknowledged entries fail closed even when the report claims
  zero live sessions: that contradiction is not safe to resolve in favour of
  the restart.
- ack_timeout_policy_applied stays strict (`value is True`), so an absent, null,
  or non-boolean value cannot open the gate on its own.

The genuine no-other-live-sessions case still passes, now justified by the
report proving sessions_live_other == 0 rather than by the absence of data.

Adds AcknowledgementFailClosedTests: 14 cases / 26 subtests covering missing,
null, empty, malformed, stale, partial-coverage, and unproven-count inputs,
the valid-acknowledgement and zero-live-session paths, timeout-policy
strictness, and that a failed check blocks proof.clean, verification, and the
restart gate.

Restart-surface suite: 132 passed, 38 subtests (branch baseline 118 passed,
12 subtests; +14 new tests, no regressions).

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01VEaP3TohHLFWkp3Z2mmuZw

drain_proof.py

@@ -0,0 +1,831 @@
+"""Pre-restart drain proof and hard gate (#661).
+
+A sanctioned MCP restart may only proceed after a machine-verifiable *drain
+proof* attests that unsafe work is clear and checkpoints are complete. Drain
+mode alone (#659) is not enough: without a proof, an apply path could still
+restart on a stale or false "ready" claim, dropping mutations and orphaning
+leases. This module defines the :class:`DrainProof` artifact, a fail-closed
+verifier, and the hard gate the sanctioned restart-apply path must consult.
+
+Design rules (mirror :mod:`restart_coordinator` / :mod:`lease_lifecycle`):
+
+* **Pure classification.** Every function here operates on already-gathered
+  inputs and returns a structured result. Nothing touches the network, the
+  filesystem, or a live process, so multi-session fixtures drive every branch.
+  This module never restarts anything; the gate only *authorizes or denies*.
+* **Fail closed.** A missing, expired, tampered, or unclean proof denies the
+  restart. Unknown checkpoint completeness is treated as *not complete*. An
+  incomplete impact report can never yield a clean proof.
+* **Non-forgeable within the process.** The proof id is a keyed hash over the
+  canonical proof contents using a per-process secret. A worker session cannot
+  hand-craft a passing proof without that secret, and a proof minted in a prior
+  daemon process will not verify after a restart (the secret is regenerated).
+* **No secrets leak.** The per-process secret never appears in a proof, an
+  ``as_dict``, an audit record, or an incident descriptor.
+
+Relationship to siblings (#655 umbrella):
+
+* **#658** ``restart_coordinator.evaluate_restart_impact`` — produces the
+  blast-radius impact report this proof consumes ("what would a restart
+  disrupt?"). ``mutations`` / ``critical_sections`` being empty is what the
+  no-in-flight-mutations check verifies.
+* **#659** graceful drain mode — performs the drain actions and calls
+  :func:`build_drain_proof` to mint the artifact once its checklist passes.
+* **#660** durable session checkpoints — supplies checkpoint completeness.
+  Because that schema may not yet be present, completeness is an *input* here,
+  never a hard table dependency; unknown fails closed.
+
+Non-goals (separate children): the emergency break-glass *workflow* (this gate
+only leaves a sanctioned bypass hole authorized elsewhere), the console UI, and
+the actual restart execution.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import hmac
+import json
+import os
+from dataclasses import dataclass
+from datetime import datetime, timedelta, timezone
+from typing import Any, Mapping, Sequence
+
+DRAIN_PROOF_VERSION = "1.0.0-issue-661"
+
+# Short default lifetime for a drain proof. A proof attests to a *point-in-time*
+# drained state; live work can resume the moment drain mode relaxes, so the
+# window in which a proof is honoured must be small (#661 security: short TTL).
+DEFAULT_PROOF_TTL_SECONDS = 120
+
+# Gate verdicts.
+GATE_ALLOW = "allow"
+GATE_DENY = "deny"
+GATE_BREAK_GLASS = "break_glass"
+
+# The mandatory drain checklist. A proof is *clean* only when every one of these
+# checks passed. Names are stable identifiers surfaced in audit + incidents.
+CHECK_NO_INFLIGHT_MUTATIONS = "no_inflight_mutations"
+CHECK_ASSIGNMENTS_STOPPED = "assignments_stopped"
+CHECK_CHECKPOINTS_COMPLETE = "checkpoints_complete"
+CHECK_HANDOFFS_OK = "handoffs_ok"
+CHECK_LEASES_HANDLED = "leases_handled"
+CHECK_ACKS_OR_TIMEOUT = "acks_or_timeout"
+
+# The only values accepted as an explicit acknowledgement from a live session.
+# Anything else — including a missing entry, a null, or a "pending"/"stale"
+# marker — leaves that session unacknowledged.
+_ACK_TOKENS = frozenset({"ack", "acked", "acknowledged"})
+
+REQUIRED_CHECKS: tuple[str, ...] = (
+    CHECK_NO_INFLIGHT_MUTATIONS,
+    CHECK_ASSIGNMENTS_STOPPED,
+    CHECK_CHECKPOINTS_COMPLETE,
+    CHECK_HANDOFFS_OK,
+    CHECK_LEASES_HANDLED,
+    CHECK_ACKS_OR_TIMEOUT,
+)
+
+
+def _utc_now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+def _parse_ts(value: str | None) -> datetime | None:
+    if not value:
+        return None
+    text = str(value).strip()
+    if not text:
+        return None
+    if text.endswith("Z"):
+        text = text[:-1] + "+00:00"
+    try:
+        dt = datetime.fromisoformat(text)
+    except ValueError:
+        return None
+    if dt.tzinfo is None:
+        dt = dt.replace(tzinfo=timezone.utc)
+    return dt
+
+
+def _canonical(payload: Any) -> str:
+    """Deterministic JSON encoding for hashing (stable key order, no spaces)."""
+
+    return json.dumps(payload, sort_keys=True, separators=(",", ":"), default=str)
+
+
+# ---------------------------------------------------------------------------
+# Per-process secret. Generated once per daemon process; regenerated on restart.
+# Injectable for tests so build + verify share a secret. Never serialized.
+# ---------------------------------------------------------------------------
+_PROCESS_SECRET = os.urandom(32)
+
+
+def process_secret() -> bytes:
+    """Return the per-process proof-signing secret (never serialized)."""
+
+    return _PROCESS_SECRET
+
+
+def _resolve_secret(secret: bytes | None) -> bytes:
+    return secret if secret is not None else _PROCESS_SECRET
+
+
+@dataclass(frozen=True)
+class DrainCheck:
+    """One mandatory drain checklist result."""
+
+    name: str
+    passed: bool
+    detail: str
+
+    def as_dict(self) -> dict[str, Any]:
+        return {"name": self.name, "passed": self.passed, "detail": self.detail}
+
+
+@dataclass(frozen=True)
+class DrainProof:
+    """Machine-verifiable proof that a restart's blast radius has been drained.
+
+    The ``proof_id`` is a keyed hash over the canonical proof body; it is the
+    tamper-evident signature verified at the gate. ``clean`` is True only when
+    every required check passed. The proof is honoured only until ``expires_at``.
+    """
+
+    version: str
+    proof_id: str
+    clean: bool
+    issued_at: str
+    expires_at: str
+    requesting_session_id: str | None
+    impact_fingerprint: str
+    checks: list[DrainCheck]
+    failed_checks: list[str]
+
+    def as_dict(self) -> dict[str, Any]:
+        return {
+            "version": self.version,
+            "proof_id": self.proof_id,
+            "clean": self.clean,
+            "issued_at": self.issued_at,
+            "expires_at": self.expires_at,
+            "requesting_session_id": self.requesting_session_id,
+            "impact_fingerprint": self.impact_fingerprint,
+            "checks": [c.as_dict() for c in self.checks],
+            "failed_checks": list(self.failed_checks),
+        }
+
+
+@dataclass(frozen=True)
+class VerifyResult:
+    """Outcome of :func:`verify_drain_proof` (fail closed)."""
+
+    valid: bool
+    reasons: list[str]
+    proof_id: str | None
+    clean: bool
+    expired: bool
+    tampered: bool
+
+    def as_dict(self) -> dict[str, Any]:
+        return {
+            "valid": self.valid,
+            "reasons": list(self.reasons),
+            "proof_id": self.proof_id,
+            "clean": self.clean,
+            "expired": self.expired,
+            "tampered": self.tampered,
+        }
+
+
+@dataclass(frozen=True)
+class GateDecision:
+    """Outcome of :func:`gate_apply_restart`."""
+
+    allow: bool
+    verdict: str
+    reasons: list[str]
+    proof_id: str | None
+    break_glass: bool
+    incident: dict[str, Any] | None
+    audit_record: dict[str, Any]
+
+    def as_dict(self) -> dict[str, Any]:
+        return {
+            "allow": self.allow,
+            "verdict": self.verdict,
+            "reasons": list(self.reasons),
+            "proof_id": self.proof_id,
+            "break_glass": self.break_glass,
+            "incident": self.incident,
+            "audit_record": dict(self.audit_record),
+        }
+
+
+def impact_fingerprint(impact_report: Mapping[str, Any] | None) -> str:
+    """Stable fingerprint of the blast-radius state a proof was minted against.
+
+    Binds a proof to the specific impact evaluation. If the live state changes
+    (a new mutation appears) between minting and gate, the caller can pass the
+    fresh fingerprint and the proof will be rejected as stale.
+    """
+
+    report = impact_report or {}
+    counts = report.get("counts") or {}
+    material = {
+        "inventory_complete": bool(report.get("inventory_complete", False)),
+        "verdict": report.get("verdict"),
+        "affected_issues": sorted(report.get("affected_issues") or []),
+        "affected_prs": sorted(report.get("affected_prs") or []),
+        "mutations": sorted(
+            str(m.get("lease_id"))
+            for m in (report.get("mutations") or [])
+            if isinstance(m, Mapping)
+        ),
+        "critical_sections": sorted(
+            str(c.get("lease_id"))
+            for c in (report.get("critical_sections") or [])
+            if isinstance(c, Mapping)
+        ),
+        "terminal_lock": bool(report.get("terminal_lock")),
+        "counts": {
+            k: counts.get(k)
+            for k in (
+                "sessions_live_other",
+                "leases_disruptive",
+                "critical_sections",
+                "mutations",
+            )
+        },
+    }
+    return hashlib.sha256(_canonical(material).encode("utf-8")).hexdigest()
+
+
+def _sign(body: Mapping[str, Any], secret: bytes) -> str:
+    """Keyed (HMAC-SHA256) signature over the canonical proof body."""
+
+    return hmac.new(
+        secret, _canonical(body).encode("utf-8"), hashlib.sha256
+    ).hexdigest()
+
+
+def _proof_body(
+    *,
+    clean: bool,
+    issued_at: str,
+    expires_at: str,
+    requesting_session_id: str | None,
+    fingerprint: str,
+    checks: Sequence[DrainCheck],
+) -> dict[str, Any]:
+    """The exact fields covered by the signature. Order-independent (canonical)."""
+
+    return {
+        "version": DRAIN_PROOF_VERSION,
+        "clean": clean,
+        "issued_at": issued_at,
+        "expires_at": expires_at,
+        "requesting_session_id": requesting_session_id,
+        "impact_fingerprint": fingerprint,
+        "checks": [c.as_dict() for c in checks],
+    }
+
+
+def _bool_input(value: Any) -> bool:
+    """Strictly interpret a drain-state flag; anything not explicitly True fails."""
+
+    return value is True
+
+
+def _live_session_count(impact_report: Mapping[str, Any]) -> int | None:
+    """Other-live-session count from the report, or ``None`` when unproven.
+
+    Only a real, non-negative integer counts. A missing ``counts`` block, a
+    malformed one, a non-integer, a bool (``True`` is an ``int`` in Python), or
+    a negative value all return ``None`` so the caller fails closed rather than
+    treating an unreadable report as "nobody was live".
+    """
+
+    counts = impact_report.get("counts")
+    if not isinstance(counts, Mapping):
+        return None
+    value = counts.get("sessions_live_other")
+    if isinstance(value, bool) or not isinstance(value, int) or value < 0:
+        return None
+    return value
+
+
+def _is_acknowledged(value: Any) -> bool:
+    """True only for an explicit acknowledgement token.
+
+    Deliberately strict: the value must already be a string carrying one of the
+    recognised tokens. Non-strings are not coerced, so ``None``, timestamps,
+    objects, and states such as "pending" or "stale" are never read as an
+    acknowledgement.
+    """
+
+    if not isinstance(value, str):
+        return False
+    return value.strip().lower() in _ACK_TOKENS
+
+
+def _evaluate_checks(
+    impact_report: Mapping[str, Any],
+    drain_state: Mapping[str, Any],
+) -> list[DrainCheck]:
+    """Compute the mandatory checklist from the impact report + drain outcomes.
+
+    The report answers "is unsafe work still in flight?"; ``drain_state`` reports
+    the drain-mode actions the coordinator/#659 performed. Every check fails
+    closed when its evidence is missing or ambiguous.
+    """
+
+    checks: list[DrainCheck] = []
+
+    inventory_complete = bool(impact_report.get("inventory_complete", False))
+    mutations = list(impact_report.get("mutations") or [])
+    critical = list(impact_report.get("critical_sections") or [])
+    disruptive = [
+        l
+        for l in (impact_report.get("affected_leases") or [])
+        if isinstance(l, Mapping) and l.get("disruptive")
+    ]
+
+    # 1. No in-flight mutations. Derived from the authoritative impact report,
+    #    not self-reported: a proof cannot claim "no mutations" while the report
+    #    still shows mutations or unsevered critical sections.
+    if not inventory_complete:
+        checks.append(
+            DrainCheck(
+                CHECK_NO_INFLIGHT_MUTATIONS,
+                False,
+                "impact report inventory incomplete; cannot confirm mutations "
+                "cleared (fail closed)",
+            )
+        )
+    elif mutations or critical:
+        checks.append(
+            DrainCheck(
+                CHECK_NO_INFLIGHT_MUTATIONS,
+                False,
+                f"{len(mutations)} mutation(s) and {len(critical)} critical "
+                "section(s) still in flight",
+            )
+        )
+    else:
+        checks.append(
+            DrainCheck(
+                CHECK_NO_INFLIGHT_MUTATIONS,
+                True,
+                "no in-flight mutations or critical sections in impact report",
+            )
+        )
+
+    # 2. New assignment halted (maintenance-drain entered).
+    checks.append(
+        DrainCheck(
+            CHECK_ASSIGNMENTS_STOPPED,
+            _bool_input(drain_state.get("assignments_stopped")),
+            "new work assignment halted"
+            if _bool_input(drain_state.get("assignments_stopped"))
+            else "assignments not confirmed stopped (fail closed)",
+        )
+    )
+
+    # 3. Durable checkpoints complete (#660). Unknown => not complete.
+    cp = drain_state.get("checkpoints_complete")
+    checks.append(
+        DrainCheck(
+            CHECK_CHECKPOINTS_COMPLETE,
+            _bool_input(cp),
+            "all live sessions checkpointed"
+            if _bool_input(cp)
+            else "checkpoint completeness unconfirmed (fail closed)",
+        )
+    )
+
+    # 4. Handoffs verified.
+    checks.append(
+        DrainCheck(
+            CHECK_HANDOFFS_OK,
+            _bool_input(drain_state.get("handoffs_verified")),
+            "pending handoffs verified"
+            if _bool_input(drain_state.get("handoffs_verified"))
+            else "handoffs not verified (fail closed)",
+        )
+    )
+
+    # 5. Leases resolved/transferred/preserved AND none left disruptive. Requires
+    #    both the drain-mode assertion and the report showing no disruptive lease.
+    leases_asserted = _bool_input(drain_state.get("leases_handled"))
+    if not leases_asserted:
+        checks.append(
+            DrainCheck(
+                CHECK_LEASES_HANDLED,
+                False,
+                "lease disposition not asserted by drain (fail closed)",
+            )
+        )
+    elif disruptive:
+        checks.append(
+            DrainCheck(
+                CHECK_LEASES_HANDLED,
+                False,
+                f"{len(disruptive)} disruptive lease(s) still active in report",
+            )
+        )
+    else:
+        checks.append(
+            DrainCheck(
+                CHECK_LEASES_HANDLED,
+                True,
+                "leases resolved/transferred/preserved; none left disruptive",
+            )
+        )
+
+    # 6. Acknowledgements received, or an explicit timeout policy was applied.
+    #
+    # Whether acknowledgements are *required* is answered by the impact report,
+    # never inferred from the shape of the drain state. Previously an absent
+    # ``acks`` key collapsed to ``{}`` and was read as "no session needed to
+    # acknowledge", so a proof minted clean — and the restart gate allowed —
+    # while the report still showed other live sessions. Absence of evidence is
+    # not evidence of absence: missing, malformed, stale, or otherwise unproven
+    # acknowledgement data fails closed whenever live sessions require
+    # acknowledgement, and only explicitly verified acknowledgement evidence
+    # may permit the operation.
+    sessions_live_other = _live_session_count(impact_report)
+    live_count_known = sessions_live_other is not None
+    # An unknown or malformed count fails closed: it cannot prove nobody had to
+    # acknowledge, so acknowledgement stays required.
+    acks_required = (not live_count_known) or sessions_live_other > 0
+
+    raw_acks = drain_state.get("acks")
+    acks_present = isinstance(raw_acks, Mapping)
+    ack_values = list(raw_acks.values()) if acks_present else []
+    acked_count = sum(1 for value in ack_values if _is_acknowledged(value))
+    every_entry_acked = bool(ack_values) and acked_count == len(ack_values)
+    covers_live_sessions = live_count_known and acked_count >= sessions_live_other
+    # Strict by construction: only an explicit ``True`` counts, so an absent,
+    # null, or non-boolean ``ack_timeout_policy_applied`` can never open this
+    # gate on its own.
+    timeout_policy = _bool_input(drain_state.get("ack_timeout_policy_applied"))
+
+    # Present-but-unacknowledged entries always fail closed, even when the
+    # report claims nobody was live: the drain state naming an outstanding
+    # session contradicts that claim, and the safe reading of a contradiction
+    # is that an acknowledgement is still owed.
+    outstanding_entries = acks_present and bool(ack_values) and not every_entry_acked
+
+    if timeout_policy and (acks_required or outstanding_entries):
+        acks_ok = True
+        detail = "explicit ack timeout policy applied and recorded"
+    elif outstanding_entries:
+        acks_ok = False
+        detail = (
+            f"{len(ack_values) - acked_count} of {len(ack_values)} "
+            "acknowledgement entries are unparseable, stale, or not "
+            "acknowledged (fail closed)"
+        )
+    elif not acks_required:
+        acks_ok = True
+        detail = (
+            "impact report proves no other live sessions required to "
+            "acknowledge (sessions_live_other=0)"
+        )
+    elif every_entry_acked and covers_live_sessions:
+        acks_ok = True
+        detail = (
+            f"{acked_count} acknowledgement entr"
+            f"{'y' if acked_count == 1 else 'ies'} verified; covers "
+            f"sessions_live_other={sessions_live_other}"
+        )
+    else:
+        acks_ok = False
+        if not live_count_known:
+            detail = (
+                "impact report does not prove the live-session count; "
+                "acknowledgement required and unproven (fail closed)"
+            )
+        elif raw_acks is None:
+            detail = (
+                "acknowledgement evidence absent while "
+                f"sessions_live_other={sessions_live_other} require "
+                "acknowledgement (fail closed)"
+            )
+        elif not acks_present:
+            detail = (
+                "acknowledgement evidence malformed: expected a mapping, got "
+                f"{type(raw_acks).__name__} (fail closed)"
+            )
+        elif not ack_values:
+            detail = (
+                "acknowledgement mapping empty while "
+                f"sessions_live_other={sessions_live_other} require "
+                "acknowledgement (fail closed)"
+            )
+        else:
+            detail = (
+                f"acknowledgements cover only {acked_count} session(s) but the "
+                f"report shows sessions_live_other={sessions_live_other} "
+                "(fail closed)"
+            )
+    checks.append(DrainCheck(CHECK_ACKS_OR_TIMEOUT, acks_ok, detail))
+
+    return checks
+
+
+def build_drain_proof(
+    *,
+    impact_report: Mapping[str, Any],
+    drain_state: Mapping[str, Any],
+    requesting_session_id: str | None = None,
+    now: datetime | None = None,
+    ttl_seconds: int = DEFAULT_PROOF_TTL_SECONDS,
+    secret: bytes | None = None,
+) -> DrainProof:
+    """Mint a drain proof from an impact report and the drain-mode outcomes.
+
+    A successful drain (every checklist item passes) yields a *clean* proof with
+    a valid signature (AC#2). An unclean drain still yields a signed proof, but
+    with ``clean=False`` and the failing checks named — the gate will deny it —
+    so the artifact is auditable rather than silently absent.
+    """
+
+    moment = now or _utc_now()
+    ttl = max(1, int(ttl_seconds))
+    issued_at = moment.isoformat()
+    expires_at = (moment + timedelta(seconds=ttl)).isoformat()
+    fingerprint = impact_fingerprint(impact_report)
+
+    checks = _evaluate_checks(impact_report, drain_state)
+    failed = [c.name for c in checks if not c.passed]
+    clean = not failed
+
+    body = _proof_body(
+        clean=clean,
+        issued_at=issued_at,
+        expires_at=expires_at,
+        requesting_session_id=requesting_session_id,
+        fingerprint=fingerprint,
+        checks=checks,
+    )
+    proof_id = _sign(body, _resolve_secret(secret))
+
+    return DrainProof(
+        version=DRAIN_PROOF_VERSION,
+        proof_id=proof_id,
+        clean=clean,
+        issued_at=issued_at,
+        expires_at=expires_at,
+        requesting_session_id=requesting_session_id,
+        impact_fingerprint=fingerprint,
+        checks=checks,
+        failed_checks=failed,
+    )
+
+
+def verify_drain_proof(
+    proof: Mapping[str, Any] | None,
+    *,
+    now: datetime | None = None,
+    secret: bytes | None = None,
+    expected_impact_fingerprint: str | None = None,
+) -> VerifyResult:
+    """Verify a drain proof, failing closed on any doubt.
+
+    A proof is valid only when: it is present and well-formed; its signature
+    recomputes with the per-process secret (not forged/tampered, not minted in a
+    prior process); it has not expired; every required check is present and
+    passed; and — when ``expected_impact_fingerprint`` is supplied — it was
+    minted against the current blast-radius state.
+    """
+
+    moment = now or _utc_now()
+    reasons: list[str] = []
+
+    if not isinstance(proof, Mapping):
+        return VerifyResult(
+            valid=False,
+            reasons=["drain proof missing or not an object (fail closed)"],
+            proof_id=None,
+            clean=False,
+            expired=False,
+            tampered=False,
+        )
+
+    proof_id = proof.get("proof_id")
+    presented_clean = bool(proof.get("clean", False))
+
+    # Rebuild the signed body from the presented fields and re-sign. Any mutation
+    # of a covered field (including flipping ``clean`` to True) breaks the match.
+    raw_checks = proof.get("checks")
+    checks: list[DrainCheck] = []
+    checks_wellformed = isinstance(raw_checks, Sequence) and not isinstance(
+        raw_checks, (str, bytes)
+    )
+    if checks_wellformed:
+        for c in raw_checks:
+            if not isinstance(c, Mapping) or "name" not in c or "passed" not in c:
+                checks_wellformed = False
+                break
+            checks.append(
+                DrainCheck(
+                    name=str(c.get("name")),
+                    passed=bool(c.get("passed")),
+                    detail=str(c.get("detail") or ""),
+                )
+            )
+
+    tampered = False
+    if not checks_wellformed:
+        reasons.append("drain proof checks malformed (fail closed)")
+        tampered = True
+    else:
+        body = _proof_body(
+            clean=presented_clean,
+            issued_at=str(proof.get("issued_at") or ""),
+            expires_at=str(proof.get("expires_at") or ""),
+            requesting_session_id=proof.get("requesting_session_id"),
+            fingerprint=str(proof.get("impact_fingerprint") or ""),
+            checks=checks,
+        )
+        expected_sig = _sign(body, _resolve_secret(secret))
+        if not (
+            isinstance(proof_id, str)
+            and hmac.compare_digest(expected_sig, proof_id)
+        ):
+            tampered = True
+            reasons.append(
+                "drain proof signature mismatch: forged, tampered, or minted "
+                "in a prior process (fail closed)"
+            )
+
+    expires = _parse_ts(proof.get("expires_at"))
+    expired = expires is None or moment >= expires
+    if expires is None:
+        reasons.append("drain proof has no valid expiry (fail closed)")
+    elif expired:
+        reasons.append(f"drain proof expired at {proof.get('expires_at')}")
+
+    # Recompute cleanliness from the checks themselves — never trust the flag.
+    recomputed_failed = [c.name for c in checks if not c.passed]
+    present_names = {c.name for c in checks}
+    missing = [name for name in REQUIRED_CHECKS if name not in present_names]
+    recomputed_clean = checks_wellformed and not recomputed_failed and not missing
+    if missing:
+        reasons.append(f"drain proof missing required checks: {', '.join(missing)}")
+    if checks_wellformed and recomputed_failed:
+        reasons.append(
+            f"drain checks failed: {', '.join(sorted(set(recomputed_failed)))}"
+        )
+    if presented_clean and not recomputed_clean:
+        tampered = True
+        reasons.append("proof claims clean but its checks do not support it")
+
+    if expected_impact_fingerprint is not None:
+        if str(proof.get("impact_fingerprint") or "") != str(
+            expected_impact_fingerprint
+        ):
+            reasons.append(
+                "drain proof was minted against a different blast-radius state "
+                "(stale; fail closed)"
+            )
+
+    valid = (not tampered) and (not expired) and recomputed_clean and not reasons
+    return VerifyResult(
+        valid=valid,
+        reasons=reasons,
+        proof_id=proof_id if isinstance(proof_id, str) else None,
+        clean=recomputed_clean,
+        expired=expired,
+        tampered=tampered,
+    )
+
+
+def _incident_descriptor(
+    *,
+    reasons: Sequence[str],
+    requesting_session_id: str | None,
+    proof_id: str | None,
+    at: str,
+) -> dict[str, Any]:
+    """Durable incident descriptor for a denied restart (caller creates issue).
+
+    Kept as data (not a live Gitea call) so this module stays pure and testable;
+    the MCP tool layer turns it into a durable issue via the incident bridge.
+    """
+
+    return {
+        "kind": "restart_drain_gate_denied",
+        "title": "Restart denied: drain proof failed the hard gate",
+        "labels": ["mcp-health", "safety", "stale-runtime", "workflow-hardening"],
+        "reasons": list(reasons),
+        "requesting_session_id": requesting_session_id,
+        "proof_id": proof_id,
+        "at": at,
+        "drain_proof_version": DRAIN_PROOF_VERSION,
+    }
+
+
+def gate_apply_restart(
+    *,
+    proof: Mapping[str, Any] | None,
+    now: datetime | None = None,
+    secret: bytes | None = None,
+    break_glass: bool = False,
+    expected_impact_fingerprint: str | None = None,
+    requesting_session_id: str | None = None,
+) -> GateDecision:
+    """Hard gate for a sanctioned restart apply (#661 AC#1/#3).
+
+    Restart apply is authorized only with a valid, unexpired, clean drain proof.
+    A missing, expired, tampered, or unclean proof denies the restart and emits
+    a durable incident descriptor. ``break_glass`` is the *only* sanctioned
+    bypass — authorization for it is the caller's responsibility (the emergency
+    workflow is a separate child); when set, the gate allows without a proof but
+    records the bypass so it is never silent.
+    """
+
+    moment = now or _utc_now()
+    at = moment.isoformat()
+
+    if break_glass:
+        reasons = ["break-glass restart authorized; drain proof gate bypassed"]
+        audit = {
+            "event": "restart_gate_evaluated",
+            "drain_proof_version": DRAIN_PROOF_VERSION,
+            "evaluated_at": at,
+            "verdict": GATE_BREAK_GLASS,
+            "allow": True,
+            "break_glass": True,
+            "requesting_session_id": requesting_session_id,
+            "proof_id": None,
+        }
+        return GateDecision(
+            allow=True,
+            verdict=GATE_BREAK_GLASS,
+            reasons=reasons,
+            proof_id=None,
+            break_glass=True,
+            incident=None,
+            audit_record=audit,
+        )
+
+    result = verify_drain_proof(
+        proof,
+        now=moment,
+        secret=secret,
+        expected_impact_fingerprint=expected_impact_fingerprint,
+    )
+
+    if result.valid:
+        reasons = ["valid unexpired clean drain proof present; restart authorized"]
+        audit = {
+            "event": "restart_gate_evaluated",
+            "drain_proof_version": DRAIN_PROOF_VERSION,
+            "evaluated_at": at,
+            "verdict": GATE_ALLOW,
+            "allow": True,
+            "break_glass": False,
+            "requesting_session_id": requesting_session_id,
+            "proof_id": result.proof_id,
+        }
+        return GateDecision(
+            allow=True,
+            verdict=GATE_ALLOW,
+            reasons=reasons,
+            proof_id=result.proof_id,
+            break_glass=False,
+            incident=None,
+            audit_record=audit,
+        )
+
+    deny_reasons = ["restart denied: drain proof invalid (fail closed)"] + list(
+        result.reasons
+    )
+    incident = _incident_descriptor(
+        reasons=deny_reasons,
+        requesting_session_id=requesting_session_id,
+        proof_id=result.proof_id,
+        at=at,
+    )
+    audit = {
+        "event": "restart_gate_evaluated",
+        "drain_proof_version": DRAIN_PROOF_VERSION,
+        "evaluated_at": at,
+        "verdict": GATE_DENY,
+        "allow": False,
+        "break_glass": False,
+        "requesting_session_id": requesting_session_id,
+        "proof_id": result.proof_id,
+        "incident_kind": incident["kind"],
+    }
+    return GateDecision(
+        allow=False,
+        verdict=GATE_DENY,
+        reasons=deny_reasons,
+        proof_id=result.proof_id,
+        break_glass=False,
+        incident=incident,
+        audit_record=audit,
+    )

gitea_mcp_server.py

@@ -2068,6 +2068,7 @@ import lease_lifecycle  # noqa: E402
 import lease_policy  # noqa: E402
 import workflow_dashboard  # noqa: E402  # #605 live queue/lease dashboard
 import restart_coordinator  # noqa: E402  # #658 MCP restart coordinator/impact
+import drain_proof  # noqa: E402  # #661 pre-restart drain proof and hard gate
 import incident_bridge  # noqa: E402
 import sentry_observability  # noqa: E402  (#606 optional Sentry observability)
 import sentry_incident_bridge  # noqa: E402  (#607 Sentry→Gitea incident bridge)
@@ -22342,6 +22343,8 @@ def gitea_request_mcp_restart(
     request_override: bool = False,
     session_id: str | None = None,
     limit: int = 200,
+    drain_proof_json: str | None = None,
+    request_break_glass: bool = False,
 ) -> dict:
     """Evaluate a proposed MCP restart and return an impact preview (#658).
 
@@ -22351,10 +22354,15 @@ def gitea_request_mcp_restart(
     verdict, so the console (#642/#652) and operators can see what a restart
     would disrupt *before* any concurrent LLM work is destroyed.
 
-    This tool is **dry-run and never restarts anything.** The mutative apply
+    This tool **never restarts a process.** In dry-run (the default) it returns
-    path is a separate child gated by a drain proof (non-goal here); calling
+    only the impact preview. With ``dry_run=False`` it enforces the #661 hard
-    with ``dry_run=False`` still performs no restart and reports that apply is
+    gate: the apply request must present a valid, unexpired, clean drain proof
-    not yet available.
+    (``drain_proof_json``) or it is denied and a durable incident descriptor is
+    returned under ``incident``. Break-glass is the only bypass and is honoured
+    only when ``request_break_glass`` is set *and* the environment carries
+    ``GITEA_BREAKGLASS_RESTART_AUTHORIZATION``. Even an authorized gate performs
+    no restart here; actual execution is a further child. The gate outcome is
+    reported under ``apply_gate`` / ``apply_authorized``.
 
     Operator override authority is read from the process environment
     (``GITEA_OPERATOR_RESTART_OVERRIDE_AUTHORIZATION``), never self-asserted by
@@ -22473,12 +22481,54 @@ def gitea_request_mcp_restart(
     payload["requesting_session_id"] = sid
     payload["operator_override_requested"] = bool(request_override)
     payload["operator_override_authorized"] = operator_authorized
+    # Actual restart execution remains a further child; this tool never restarts
+    # a process. What #661 adds is the *hard gate*: an apply request (dry_run
+    # False) must present a valid, unexpired, clean drain proof, or it is denied
+    # and a durable incident is raised. Break-glass is the only bypass and its
+    # authorization is read from the environment, never self-asserted.
     payload["apply_supported"] = False
     if not dry_run:
-        payload["reasons"] = list(payload.get("reasons") or []) + [
+        proof_obj: dict | None = None
-            "apply requested but not supported: sanctioned restart apply is "
+        proof_parse_error: str | None = None
-            "gated by a drain proof (separate child); no restart performed (#658)"
+        if drain_proof_json:
-        ]
+            try:
+                parsed = json.loads(drain_proof_json)
+                proof_obj = parsed if isinstance(parsed, dict) else None
+                if proof_obj is None:
+                    proof_parse_error = "drain_proof_json is not a JSON object"
+            except (ValueError, TypeError) as exc:
+                proof_parse_error = f"invalid drain_proof_json: {_redact(str(exc))}"
+
+        break_glass_authorized = bool(
+            (
+                os.environ.get("GITEA_BREAKGLASS_RESTART_AUTHORIZATION") or ""
+            ).strip()
+        )
+        break_glass = bool(request_break_glass and break_glass_authorized)
+
+        expected_fp = drain_proof.impact_fingerprint(report.as_dict())
+        gate = drain_proof.gate_apply_restart(
+            proof=proof_obj,
+            break_glass=break_glass,
+            expected_impact_fingerprint=expected_fp,
+            requesting_session_id=sid,
+        )
+        gate_payload = gate.as_dict()
+        if proof_parse_error and not break_glass:
+            gate_payload["reasons"] = [proof_parse_error] + list(
+                gate_payload.get("reasons") or []
+            )
+        payload["apply_gate"] = gate_payload
+        payload["apply_authorized"] = gate.allow
+        payload["break_glass_requested"] = bool(request_break_glass)
+        payload["break_glass_authorized"] = break_glass_authorized
+        # Even an authorized gate performs no restart here: execution is a later
+        # child. The gate proves the apply path *would* be permitted.
+        payload["reasons"] = list(payload.get("reasons") or []) + list(
+            gate_payload.get("reasons") or []
+        )
+        if not gate.allow and gate.incident is not None:
+            payload["incident"] = gate.incident
     return payload
 
 

tests/test_drain_proof.py

@@ -0,0 +1,596 @@
+"""Tests for the pre-restart drain proof and hard gate (#661).
+
+Covers the acceptance criteria:
+
+1. Restart apply without a proof fails closed.
+2. A successful drain produces a verifiable proof.
+3. An open unsafe mutation makes the proof fail (multi-session fixture).
+4. Pass / fail / expired verification paths.
+
+Plus the security posture: forged/tampered proofs are rejected, break-glass is
+the only bypass and is never silent, a stale blast-radius fingerprint rejects a
+proof, and no per-process secret ever leaks into a serialized artifact.
+"""
+
+from __future__ import annotations
+
+import os
+import unittest
+from datetime import datetime, timedelta, timezone
+
+import drain_proof as dp
+import restart_coordinator as rc
+
+
+NOW = datetime(2026, 7, 24, 6, 0, 0, tzinfo=timezone.utc)
+SECRET = b"unit-test-drain-proof-secret-0123456789abcdef"
+
+
+def _live_pid() -> int:
+    return os.getpid()
+
+
+def _clean_drain_state() -> dict:
+    """Every drain action succeeded, no sessions outstanding."""
+
+    return {
+        "assignments_stopped": True,
+        "checkpoints_complete": True,
+        "handoffs_verified": True,
+        "leases_handled": True,
+        "acks": {},  # no other live sessions to acknowledge
+        "ack_timeout_policy_applied": False,
+    }
+
+
+def _safe_report() -> dict:
+    """Impact report with no other live work: a restart here is safe."""
+
+    report = rc.evaluate_restart_impact(
+        {"sessions": [], "leases": [], "inventory_complete": True},
+        now=NOW,
+        requesting_session_id="prgs-controller-1-req",
+    )
+    return report.as_dict()
+
+
+def _unsafe_mutation_report() -> dict:
+    """Multi-session report: a second session holds a live author mutation."""
+
+    sessions = [
+        {
+            "session_id": "prgs-controller-1-req",
+            "role": "controller",
+            "profile": "prgs-controller",
+            "pid": _live_pid(),
+            "status": "active",
+            "last_heartbeat_at": NOW.isoformat(),
+        },
+        {
+            "session_id": "prgs-author-99",
+            "role": "author",
+            "profile": "prgs-author",
+            "pid": _live_pid(),
+            "status": "active",
+            "last_heartbeat_at": NOW.isoformat(),
+        },
+    ]
+    leases = [
+        {
+            "lease_id": "lease-mut",
+            "session_id": "prgs-author-99",
+            "role": "author",
+            "phase": "implementing",
+            "work_kind": "issue",
+            "work_number": 661,
+            "worktree_path": "branches/issue-661",
+            "freshness": {"freshness": "active"},
+        }
+    ]
+    report = rc.evaluate_restart_impact(
+        {"sessions": sessions, "leases": leases, "inventory_complete": True},
+        now=NOW,
+        requesting_session_id="prgs-controller-1-req",
+    )
+    return report.as_dict()
+
+
+class BuildDrainProofTests(unittest.TestCase):
+    def test_clean_drain_produces_verifiable_clean_proof(self):
+        """AC#2: a successful drain produces a verifiable proof."""
+
+        proof = dp.build_drain_proof(
+            impact_report=_safe_report(),
+            drain_state=_clean_drain_state(),
+            requesting_session_id="prgs-controller-1-req",
+            now=NOW,
+            secret=SECRET,
+        )
+        self.assertTrue(proof.clean)
+        self.assertEqual(proof.failed_checks, [])
+        self.assertEqual(
+            {c.name for c in proof.checks}, set(dp.REQUIRED_CHECKS)
+        )
+        result = dp.verify_drain_proof(
+            proof.as_dict(), now=NOW, secret=SECRET
+        )
+        self.assertTrue(result.valid, result.reasons)
+        self.assertFalse(result.expired)
+        self.assertFalse(result.tampered)
+
+    def test_open_mutation_makes_proof_unclean(self):
+        """AC#3: an unsafe mutation still in flight fails the proof."""
+
+        proof = dp.build_drain_proof(
+            impact_report=_unsafe_mutation_report(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        )
+        self.assertFalse(proof.clean)
+        self.assertIn(dp.CHECK_NO_INFLIGHT_MUTATIONS, proof.failed_checks)
+        # Leases-handled also fails: the report still shows a disruptive lease.
+        self.assertIn(dp.CHECK_LEASES_HANDLED, proof.failed_checks)
+        result = dp.verify_drain_proof(proof.as_dict(), now=NOW, secret=SECRET)
+        self.assertFalse(result.valid)
+
+    def test_incomplete_inventory_fails_no_mutations_check(self):
+        proof = dp.build_drain_proof(
+            impact_report={"inventory_complete": False},
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        )
+        self.assertFalse(proof.clean)
+        self.assertIn(dp.CHECK_NO_INFLIGHT_MUTATIONS, proof.failed_checks)
+
+    def test_missing_checkpoint_flag_fails_closed(self):
+        state = _clean_drain_state()
+        del state["checkpoints_complete"]
+        proof = dp.build_drain_proof(
+            impact_report=_safe_report(), drain_state=state, now=NOW, secret=SECRET
+        )
+        self.assertFalse(proof.clean)
+        self.assertIn(dp.CHECK_CHECKPOINTS_COMPLETE, proof.failed_checks)
+
+    def test_non_true_flags_fail_closed(self):
+        """A truthy-but-not-True value (e.g. the string 'yes') must not pass."""
+
+        state = _clean_drain_state()
+        state["assignments_stopped"] = "yes"
+        proof = dp.build_drain_proof(
+            impact_report=_safe_report(), drain_state=state, now=NOW, secret=SECRET
+        )
+        self.assertIn(dp.CHECK_ASSIGNMENTS_STOPPED, proof.failed_checks)
+
+    def test_ack_timeout_policy_satisfies_ack_check(self):
+        state = _clean_drain_state()
+        state["acks"] = {"prgs-author-99": "pending"}
+        state["ack_timeout_policy_applied"] = True
+        proof = dp.build_drain_proof(
+            impact_report=_safe_report(), drain_state=state, now=NOW, secret=SECRET
+        )
+        names = {c.name: c.passed for c in proof.checks}
+        self.assertTrue(names[dp.CHECK_ACKS_OR_TIMEOUT])
+
+    def test_outstanding_acks_without_timeout_fail(self):
+        state = _clean_drain_state()
+        state["acks"] = {"prgs-author-99": "pending"}
+        state["ack_timeout_policy_applied"] = False
+        proof = dp.build_drain_proof(
+            impact_report=_safe_report(), drain_state=state, now=NOW, secret=SECRET
+        )
+        self.assertIn(dp.CHECK_ACKS_OR_TIMEOUT, proof.failed_checks)
+
+    def test_all_acked_satisfies_ack_check(self):
+        state = _clean_drain_state()
+        state["acks"] = {"prgs-author-99": "acked", "prgs-author-2": "acknowledged"}
+        proof = dp.build_drain_proof(
+            impact_report=_safe_report(), drain_state=state, now=NOW, secret=SECRET
+        )
+        names = {c.name: c.passed for c in proof.checks}
+        self.assertTrue(names[dp.CHECK_ACKS_OR_TIMEOUT])
+
+
+class VerifyDrainProofTests(unittest.TestCase):
+    def _clean_proof_dict(self) -> dict:
+        return dp.build_drain_proof(
+            impact_report=_safe_report(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        ).as_dict()
+
+    def test_missing_proof_is_invalid(self):
+        result = dp.verify_drain_proof(None, now=NOW, secret=SECRET)
+        self.assertFalse(result.valid)
+        self.assertIsNone(result.proof_id)
+
+    def test_expired_proof_is_invalid(self):
+        """AC#4: an expired proof fails verification."""
+
+        proof = self._clean_proof_dict()
+        later = NOW + timedelta(seconds=dp.DEFAULT_PROOF_TTL_SECONDS + 1)
+        result = dp.verify_drain_proof(proof, now=later, secret=SECRET)
+        self.assertFalse(result.valid)
+        self.assertTrue(result.expired)
+
+    def test_proof_valid_just_before_expiry(self):
+        proof = self._clean_proof_dict()
+        almost = NOW + timedelta(seconds=dp.DEFAULT_PROOF_TTL_SECONDS - 1)
+        result = dp.verify_drain_proof(proof, now=almost, secret=SECRET)
+        self.assertTrue(result.valid, result.reasons)
+
+    def test_wrong_secret_rejected(self):
+        """A proof minted in a prior process (different secret) will not verify."""
+
+        proof = self._clean_proof_dict()
+        result = dp.verify_drain_proof(proof, now=NOW, secret=b"other-secret")
+        self.assertFalse(result.valid)
+        self.assertTrue(result.tampered)
+
+    def test_flipping_clean_flag_is_detected(self):
+        """Forging clean=True on an unclean proof breaks the signature."""
+
+        unclean = dp.build_drain_proof(
+            impact_report=_unsafe_mutation_report(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        ).as_dict()
+        self.assertFalse(unclean["clean"])
+        unclean["clean"] = True  # forge
+        result = dp.verify_drain_proof(unclean, now=NOW, secret=SECRET)
+        self.assertFalse(result.valid)
+        self.assertTrue(result.tampered)
+
+    def test_tampering_a_check_is_detected(self):
+        unclean = dp.build_drain_proof(
+            impact_report=_unsafe_mutation_report(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        ).as_dict()
+        for c in unclean["checks"]:
+            if c["name"] == dp.CHECK_NO_INFLIGHT_MUTATIONS:
+                c["passed"] = True  # forge the failing check to pass
+        result = dp.verify_drain_proof(unclean, now=NOW, secret=SECRET)
+        self.assertFalse(result.valid)
+        self.assertTrue(result.tampered)
+
+    def test_missing_required_check_rejected(self):
+        proof = self._clean_proof_dict()
+        proof["checks"] = [
+            c for c in proof["checks"] if c["name"] != dp.CHECK_HANDOFFS_OK
+        ]
+        result = dp.verify_drain_proof(proof, now=NOW, secret=SECRET)
+        self.assertFalse(result.valid)
+
+    def test_stale_fingerprint_rejected(self):
+        proof = self._clean_proof_dict()
+        result = dp.verify_drain_proof(
+            proof,
+            now=NOW,
+            secret=SECRET,
+            expected_impact_fingerprint="deadbeef",
+        )
+        self.assertFalse(result.valid)
+
+    def test_matching_fingerprint_accepted(self):
+        report = _safe_report()
+        proof = dp.build_drain_proof(
+            impact_report=report,
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        ).as_dict()
+        fp = dp.impact_fingerprint(report)
+        result = dp.verify_drain_proof(
+            proof, now=NOW, secret=SECRET, expected_impact_fingerprint=fp
+        )
+        self.assertTrue(result.valid, result.reasons)
+
+
+class GateApplyRestartTests(unittest.TestCase):
+    def _clean_proof_dict(self) -> dict:
+        return dp.build_drain_proof(
+            impact_report=_safe_report(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        ).as_dict()
+
+    def test_apply_without_proof_denied(self):
+        """AC#1: restart apply without a proof fails closed + raises incident."""
+
+        decision = dp.gate_apply_restart(proof=None, now=NOW, secret=SECRET)
+        self.assertFalse(decision.allow)
+        self.assertEqual(decision.verdict, dp.GATE_DENY)
+        self.assertIsNotNone(decision.incident)
+        self.assertEqual(
+            decision.incident["kind"], "restart_drain_gate_denied"
+        )
+
+    def test_apply_with_valid_proof_allowed(self):
+        decision = dp.gate_apply_restart(
+            proof=self._clean_proof_dict(), now=NOW, secret=SECRET
+        )
+        self.assertTrue(decision.allow)
+        self.assertEqual(decision.verdict, dp.GATE_ALLOW)
+        self.assertIsNone(decision.incident)
+
+    def test_apply_with_expired_proof_denied_with_incident(self):
+        later = NOW + timedelta(seconds=dp.DEFAULT_PROOF_TTL_SECONDS + 5)
+        decision = dp.gate_apply_restart(
+            proof=self._clean_proof_dict(), now=later, secret=SECRET
+        )
+        self.assertFalse(decision.allow)
+        self.assertIsNotNone(decision.incident)
+
+    def test_apply_with_unclean_proof_denied(self):
+        """AC#3 at the gate: an unsafe-mutation proof is denied."""
+
+        unclean = dp.build_drain_proof(
+            impact_report=_unsafe_mutation_report(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        ).as_dict()
+        decision = dp.gate_apply_restart(proof=unclean, now=NOW, secret=SECRET)
+        self.assertFalse(decision.allow)
+        self.assertIsNotNone(decision.incident)
+
+    def test_break_glass_allows_without_proof_but_records_bypass(self):
+        decision = dp.gate_apply_restart(
+            proof=None, now=NOW, secret=SECRET, break_glass=True
+        )
+        self.assertTrue(decision.allow)
+        self.assertEqual(decision.verdict, dp.GATE_BREAK_GLASS)
+        self.assertTrue(decision.break_glass)
+        self.assertIsNone(decision.incident)
+        self.assertTrue(decision.audit_record["break_glass"])
+
+    def test_denied_gate_carries_stale_fingerprint_reason(self):
+        decision = dp.gate_apply_restart(
+            proof=self._clean_proof_dict(),
+            now=NOW,
+            secret=SECRET,
+            expected_impact_fingerprint="not-the-fingerprint",
+        )
+        self.assertFalse(decision.allow)
+
+
+class SecretHygieneTests(unittest.TestCase):
+    def test_secret_never_serialized(self):
+        proof = dp.build_drain_proof(
+            impact_report=_safe_report(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+            secret=SECRET,
+        )
+        blob = dp._canonical(proof.as_dict())
+        self.assertNotIn(SECRET.decode(), blob)
+        # The signature is a hex digest, not the raw secret.
+        self.assertNotIn(SECRET.hex(), blob)
+
+    def test_incident_descriptor_has_no_secret(self):
+        decision = dp.gate_apply_restart(proof=None, now=NOW, secret=SECRET)
+        blob = dp._canonical(decision.incident)
+        self.assertNotIn(SECRET.decode(), blob)
+
+
+def _drained_report_with_live_sessions(count: int) -> dict:
+    """Report with ``count`` other live sessions but nothing in flight.
+
+    Every other checklist item passes against this report, so a failure
+    isolates the acknowledgement check rather than tripping on mutations.
+    """
+
+    sessions = [
+        {
+            "session_id": "prgs-controller-1-req",
+            "role": "controller",
+            "profile": "prgs-controller",
+            "pid": _live_pid(),
+            "status": "active",
+            "last_heartbeat_at": NOW.isoformat(),
+        }
+    ]
+    for index in range(count):
+        sessions.append(
+            {
+                "session_id": f"prgs-author-{index}",
+                "role": "author",
+                "profile": "prgs-author",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            }
+        )
+    report = rc.evaluate_restart_impact(
+        {"sessions": sessions, "leases": [], "inventory_complete": True},
+        now=NOW,
+        requesting_session_id="prgs-controller-1-req",
+    )
+    return report.as_dict()
+
+
+class AcknowledgementFailClosedTests(unittest.TestCase):
+    """Acknowledgement evidence must fail closed unless explicitly verified.
+
+    Regression cover for the reviewed fail-open on PR #882: an absent ``acks``
+    key collapsed to ``{}`` and was read as "no other live sessions required to
+    acknowledge", so a proof minted clean and the restart gate allowed while the
+    impact report still showed other live sessions.
+    """
+
+    def _state(self, **overrides) -> dict:
+        state = _clean_drain_state()
+        state.pop("acks", None)
+        state["ack_timeout_policy_applied"] = False
+        state.update(overrides)
+        return state
+
+    def _acks_check(self, proof) -> dp.DrainCheck:
+        return next(c for c in proof.checks if c.name == dp.CHECK_ACKS_OR_TIMEOUT)
+
+    def _build(self, report: dict, state: dict):
+        return dp.build_drain_proof(
+            impact_report=report, drain_state=state, now=NOW, secret=SECRET
+        )
+
+    def assertAcksFailClosed(self, report: dict, state: dict) -> None:
+        proof = self._build(report, state)
+        self.assertFalse(self._acks_check(proof).passed)
+        self.assertIn(dp.CHECK_ACKS_OR_TIMEOUT, proof.failed_checks)
+        self.assertFalse(proof.clean)
+
+    # --- missing / null / empty / malformed ------------------------------
+
+    def test_missing_acks_key_with_live_sessions_fails_closed(self):
+        """The exact reviewed defect: absent key, three other live sessions."""
+        report = _drained_report_with_live_sessions(3)
+        self.assertEqual(report["counts"]["sessions_live_other"], 3)
+        state = self._state()
+        self.assertNotIn("acks", state)
+        proof = self._build(report, state)
+        check = self._acks_check(proof)
+        self.assertFalse(check.passed)
+        self.assertNotIn("no other live sessions", check.detail)
+        self.assertIn("fail closed", check.detail)
+        self.assertFalse(proof.clean)
+        self.assertEqual(proof.failed_checks, [dp.CHECK_ACKS_OR_TIMEOUT])
+
+    def test_none_acks_with_live_sessions_fails_closed(self):
+        self.assertAcksFailClosed(
+            _drained_report_with_live_sessions(2), self._state(acks=None)
+        )
+
+    def test_empty_acks_with_live_sessions_fails_closed(self):
+        self.assertAcksFailClosed(
+            _drained_report_with_live_sessions(1), self._state(acks={})
+        )
+
+    def test_malformed_acks_fail_closed(self):
+        for malformed in ([], "ack", 7, ("ack",), True):
+            with self.subTest(malformed=malformed):
+                self.assertAcksFailClosed(
+                    _drained_report_with_live_sessions(1),
+                    self._state(acks=malformed),
+                )
+
+    # --- stale / unproven values -----------------------------------------
+
+    def test_stale_or_unproven_ack_values_fail_closed(self):
+        for value in ("pending", "stale", "unknown", "", None, True, 1, NOW):
+            with self.subTest(value=value):
+                self.assertAcksFailClosed(
+                    _drained_report_with_live_sessions(1),
+                    self._state(acks={"prgs-author-0": value}),
+                )
+
+    def test_partial_coverage_fails_closed(self):
+        """Fewer acknowledgements than the report's live-session count."""
+        self.assertAcksFailClosed(
+            _drained_report_with_live_sessions(3),
+            self._state(acks={"prgs-author-0": "ack"}),
+        )
+
+    def test_one_unacked_entry_among_many_fails_closed(self):
+        self.assertAcksFailClosed(
+            _drained_report_with_live_sessions(2),
+            self._state(acks={"prgs-author-0": "ack", "prgs-author-1": "pending"}),
+        )
+
+    def test_unproven_live_session_count_fails_closed(self):
+        """A missing or malformed count cannot prove nobody had to acknowledge."""
+        malformed_counts = (
+            None,
+            {},
+            {"sessions_live_other": None},
+            {"sessions_live_other": "3"},
+            {"sessions_live_other": -1},
+            {"sessions_live_other": True},
+        )
+        for counts in malformed_counts:
+            with self.subTest(counts=counts):
+                report = _drained_report_with_live_sessions(0)
+                if counts is None:
+                    report.pop("counts", None)
+                else:
+                    report["counts"] = counts
+                self.assertAcksFailClosed(report, self._state())
+
+    # --- valid evidence still passes -------------------------------------
+
+    def test_complete_valid_acks_pass(self):
+        report = _drained_report_with_live_sessions(2)
+        state = self._state(
+            acks={"prgs-author-0": "ack", "prgs-author-1": "acknowledged"}
+        )
+        proof = self._build(report, state)
+        self.assertTrue(self._acks_check(proof).passed)
+        self.assertTrue(proof.clean)
+        self.assertEqual(proof.failed_checks, [])
+
+    def test_no_other_live_sessions_still_passes(self):
+        """Intended behavior retained: zero live sessions needs no acks."""
+        report = _drained_report_with_live_sessions(0)
+        self.assertEqual(report["counts"]["sessions_live_other"], 0)
+        proof = self._build(report, self._state())
+        check = self._acks_check(proof)
+        self.assertTrue(check.passed)
+        self.assertIn("sessions_live_other=0", check.detail)
+        self.assertTrue(proof.clean)
+
+    # --- timeout policy cannot become a second fail-open ------------------
+
+    def test_unproven_timeout_policy_cannot_open_the_gate(self):
+        for value in (None, "true", "yes", 1, "True", [], {}):
+            with self.subTest(value=value):
+                self.assertAcksFailClosed(
+                    _drained_report_with_live_sessions(2),
+                    self._state(ack_timeout_policy_applied=value),
+                )
+
+    def test_explicit_timeout_policy_permits(self):
+        proof = self._build(
+            _drained_report_with_live_sessions(2),
+            self._state(ack_timeout_policy_applied=True),
+        )
+        check = self._acks_check(proof)
+        self.assertTrue(check.passed)
+        self.assertIn("timeout policy", check.detail)
+        self.assertTrue(proof.clean)
+
+    # --- the gate itself must deny ---------------------------------------
+
+    def test_failed_ack_check_denies_the_restart_gate(self):
+        report = _drained_report_with_live_sessions(3)
+        proof = self._build(report, self._state())
+        self.assertFalse(proof.clean)
+        decision = dp.gate_apply_restart(
+            proof=proof.as_dict(),
+            now=NOW,
+            secret=SECRET,
+            expected_impact_fingerprint=dp.impact_fingerprint(report),
+        )
+        self.assertFalse(decision.allow)
+        self.assertEqual(decision.verdict, dp.GATE_DENY)
+        self.assertIsNotNone(decision.incident)
+
+    def test_unclean_ack_proof_fails_verification(self):
+        report = _drained_report_with_live_sessions(3)
+        proof = self._build(report, self._state())
+        result = dp.verify_drain_proof(
+            proof.as_dict(),
+            now=NOW,
+            secret=SECRET,
+            expected_impact_fingerprint=dp.impact_fingerprint(report),
+        )
+        self.assertFalse(result.valid)
+        self.assertFalse(result.clean)
+
+
+if __name__ == "__main__":
+    unittest.main()