PR Review Summary

Status: Approved

Night-shift marker per guide §0: operator-routed primary review (gpt out until 2:30am); per the current roster reading my APPROVED clears the §6.1 gate — qualification remains @tobiu's merge call.

🪜 Strategic-Fit Decision

• Decision: Approve
• Rationale: The fix reconciles on the correct axis (already-woken, not merely unread), the pure-module split matches the established wake-daemon decomposition, the conservative-keep discipline protects the never-withhold-a-genuine-wake invariant, and the one real exposure I hunted for is strictly dominated by an already-ticketed sibling failure (detail in Depth Floor). Both findings are non-blocking with concrete follow-through.

Peer-Review Opening: This kills a failure class I have unusually direct standing on — my own session received 8+ stale doorbell redeliveries tonight, several reporting "N events" for messages long processed. The digest that cried HIGH dies with this merge.

🧭 Patch-Blind Premise Snapshot

• Inputs Read Before Patch: #12850's premise (the #12479 residual: readAt is the wrong reconcile axis); the wake daemon source read at length tonight (coalesce state, WAKE_PRIORITY_RANKS, flushSubscription, the STATE_FILE/LOG_FILE daemon-state pattern); the wake-redelivery doctrine (at-least-once doorbells, mailbox = truth); the fresh #12855 lineage (cursor semantics + the atomic-write precedent in the same daemon); my own session's stale-doorbell empirics.
• Expected Solution Shape: a per-subscription durable high-water-mark of already-woken-for logIds consumed by flushSubscription to bound digest count/priority — pure logic in a wake/ sibling module, durable in DAEMON_DATA_DIR, composing with (not replacing) the readAt reconcile. Two boundaries it must NOT cross: never suppress a genuine wake (malformed events must fail open), and the watermark must not introduce a new permanent-silence mode on logId regression.
• Patch Verdict: Matches and improves. Improvements over my expected shape: the Number(null) === 0 coercion trap is explicitly handled with the null-check-precedes-coercion comment (a subtle real bug pre-empted); maxLogId returns null — not 0 — for logId-less sets so the caller never resets the watermark downward; the advance is forward-only by explicit comparison. The suppression semantics are doctrine-consistent: already-woken means the doorbell already fired once; the mailbox holds the truth — same contract the readAt suppress established.

🕸️ Context & Graph Linking

• Target Epic / Issue ID: Resolves #12850
• Related Graph Nodes: #12479 (hypothesis-1 lineage, honestly cited), #12855/#12849 (sibling cursor semantics), #12862 (the regression class — see Depth Floor), flushDeferPolicy.mjs (the composing defer)

🔬 Depth Floor

Challenge 1 (non-blocking — dominated by an already-ticketed sibling, with follow-through): the durable watermark inherits the logId-regression silence class: after a GraphLog reset/restore with a surviving DAEMON_DATA_DIR, a stale high watermark sits above every new event forever → permanent digest silence for that subscription. I carried this question in deliberately (it is the watermark-shaped twin of the cursor-ahead-of-MAX hypothesis from my #12855 review) — and verified it is strictly dominated: in the identical scenario the upstream lastSyncId poll cursor (WHERE log_id > <stale-high>) goes silent before any event reaches flushSubscription, so this PR adds no new reachable silence. The class is already ticketed as #12862 (boardless) — I am extending that ticket's scope to clamp BOTH cursors (poll cursor + woken watermark) so the eventual fix covers them together; a one-line code comment acknowledging the shared reset-class dependency would be welcome but is not required.

Challenge 2 (non-blocking nit — consistency with your own fresh precedent): persistWokenWatermark uses bare fs.writeFileSync — the exact kill-truncation pattern your #12855 replaced with temp+renameSync for lastSyncId in this same daemon, hours ago. Here the blast radius is genuinely different: loadWokenWatermark is corrupt-tolerant and a torn file degrades to at most one cycle of re-counted backlog (the bug's own symptom, once, self-healing) versus the cursor's full-backlog flood — so by-design tolerance makes this a consistency nit, not a defect. Worth a 3-line tighten whenever the file is next touched.

Rhetorical-Drift Audit (per guide §7.4):

• PR description: every mechanical claim verified against the diff — all-four-event-types filtering ✓ (post-readAt, pre-suppress), forward-only advance ✓, durable + corrupt-tolerant load ✓, no config-template surface ✓ (the no-leaf delta is argued correctly: daemon-internal state mirrors STATE_FILE, only cross-process state earns leaves)
• The "composes with — does not replace" framing is mechanically true (the readAt filter remains line-for-line; the defer policy untouched)
• The #12479 hypothesis-1 citation is honest lineage, not borrowed authority
• Evidence line: L2 claimed, L3 residuals explicitly listed post-merge — no evidence-class collapse

Findings: Pass.

🧠 Graph Ingestion Notes

• [RETROSPECTIVE]: "Reconcile on the right axis" is the durable lesson: readAt tracks consumption, the watermark tracks notification — conflating them produced a bug class that survived one fix (#12479) because the axis, not the implementation, was wrong. Pairs with tonight's ticket-create lesson (read-status ≠ relevance) — two instances, same night, of state-axis confusion.
• [TOOLING_GAP]: fresh review worktrees cannot run the daemon-spawn specs — the spawned child crashes importing the absent gitignored config.mjs, failing 21/70 with misleading delivery-timeout errors. Materializing configs into the worktree yields the true result. Reviewers checking out daemon PRs need this step (the #12857 bootstrapWorktree fix covers normal worktrees; manual git worktree add skips it).

N/A Audits — 🎯 📑 🪜 📡 🛂 🔌

N/A beyond items covered above: close-target Resolves #12850 verified leaf + newline-isolated + commit bodies grepped clean; daemon-internal state (no public contract → no ledger; Evidence audited in §7.4); no OpenAPI; established pattern extension (no novel abstraction); the watermark file is new daemon-private state, not a consumed wire format.

🧪 Test-Execution & Location Audit

• Branch checked out at head 38c7f1c6c in an isolated worktree
• Ran the full wake unit dir myself: 70/70 passed (16.3s) after materializing configs — independently reproducing the author's count (first run failed 21 daemon-spawn specs purely from the worktree config gap; see [TOOLING_GAP])
• New spec location canonical (test/playwright/unit/ai/daemons/wake/); the 6 falsifiers include the AC2 spoof repro and the all-types AC4 regression
• CI green on current head

Findings: Tests pass — independently verified, with the env-artifact explicitly falsified rather than reported as failure.

📋 Required Actions

No required actions — eligible for human merge.

📊 Evaluation Metrics

• [ARCH_ALIGNMENT]: 100 - I actively considered the pure-module decomposition convention (matches flushDeferPolicy exactly), config-surface discipline (daemon-internal state correctly kept off the template), doctrine consistency (at-least-once + mailbox-truth preserved), and compose-vs-replace layering (verified line-level) — none violated.
• [CONTENT_COMPLETENESS]: 100 - I checked for undocumented helpers (both carry full Anchor & Echo with the coercion-trap rationale), missing PR-body sections, and unexplained deltas (both deltas argued with substance) — none apply.
• [EXECUTION_QUALITY]: 95 - 5 points deducted for the non-atomic watermark persist (Challenge 2): self-healing by design, but inconsistent with the atomic-write discipline this same daemon established hours earlier.
• [PRODUCTIVITY]: 100 - All four ticket ACs delivered; the AC2 spoof repro and AC4 never-drop regression are precisely the two falsifiers that matter.
• [IMPACT]: 60 - Substantive swarm-operations fix: every maintainer's wake quality improves, and the spoofed-HIGH class that erodes trust in priority signals dies.
• [COMPLEXITY]: 30 - Low-moderate: one pure module + flush wiring + persistence; the depth is semantic (axis choice, coercion traps, monotonicity), not structural.
• [EFFORT_PROFILE]: Quick Win - High swarm-wide ROI at low structural complexity, with the failure class empirically live in tonight's own session.

Eligible for human merge — and tonight's stale-doorbell noise in my own transcript is the before-picture this PR retires.