[entelecheia]

Refs #367.

The SessionStart stage probe now validates all 32 literal anchors across six current sources before measurement and checks a separate runtime error receipt after every invocation. Missing or duplicate anchors name the anchor and source file and fail the step even when bin/amem.js catches the import error. Stale notice/context and worker anchors are updated. Cheap unit tests enumerate every anchor, compile the transformed sources and prove that a zero-exit caller cannot hide an anchor failure. The existing five cold attempts, full-foreign diagnostic and timing policy are retained.

A separate reported-only transcript runner observes the same Stop, SessionEnd and 100-title StopBurst workload shape, preserving the published executable's real clock and deadlines. src/ and scripts/measure-transcript-budget.ts are byte-identical to main. The strict transcript command remains the gate. The new CI step runs on both hosted OSes after the strict performance steps, with its own reported-only result and observer-cost summary. It adds no budget advisory, retry or limit change.

Numeric phase receipts cover bootstrap/import, worker spawn and entry, metadata, scan and cursor open/commit/close, first replay, CaptureStore open/drain/transactions/close, stop/live-RSS/kill/reap/output/abandon, accounting/final RSS, variant append, metric sampling/write and the exit event. Parent and worker timestamps bracket first-drain-to-metric directly, including the 254-262 ms interval from the Linux evidence. CPU user/system endpoints and matched first/last interval deltas are retained. Missing or count-mismatched intervals remain null.

Observation is bounded: two paired logical runs per event, at most 64 resumptions, a cooperative two-minute whole-run cap, eight processes per invocation, 4096 marks/process, and 64 fixed 512-byte numeric slots/process. Each phase retains first/last/count/CPU values. Completed slot writes remain readable after SIGKILL; malformed or torn receipts fail instead of inventing a complete trace (there is no claim of atomic interrupted overwrite). Every invocation has a concise receipt; full details retain each first observed call and the eight largest observed wall/drain intervals. Truncation and incomplete coverage are explicit.

Within this separate diagnostic, uninstrumented and observed arms alternate order and retain the strict harness's warm-home policy and fresh-home StopBurst policy. Inline recording/loader costs are labeled lower bounds, with paired first-call wall deltas labeled noncausal because deadline-dependent completed work differs. No delta is subtracted from any strict result. Observed cleanup uses recorded worker PIDs/groups; uninstrumented helper-command census is labeled separately rather than fabricating PID coverage. Surviving/unknown owned cleanup retains its fixture and fails the diagnostic.

Validation:

• Remote Bun 1.4.2 at 9f2ec81: typecheck, boundary (172 source files), 10 focused tests / 362 assertions passed. Repaired stage probe reached known work (cold attempt0 completed72.30ms) and its separate full-foreign case. Unit controls include a killed observer, malformed/torn slots, source drift hidden by a zero-exit bin, actual native bootstrap and an injected270ms worker-close stall correctly bracketed by >=250ms receipts.
• Coverage follow-up e986e8d: bounded real runner exited0 with 43 total calls; all25 observed calls captured worker entry, scan, first drain and matching metric; no truncated traces; control and observed each completed6/6 logical runs. Output292573bytes/13details. Measured record+loader lower bounds1.98-3.46ms/call (median2.43); paired first-call deltas-1.57 to13.82ms. This is fixture/observer evidence, not hosted acceptance or an attribution of the historical delay.
• Final head e986e8db74531ca659b6588d80aa6f46174b8be3, remote Bun 1.4.2: typecheck, boundary (172 source files) and pack dry-run (182 files) passed. Full suite: 3,688 pass, 26 existing platform skips, 0 fail; 50,817 assertions, 3,714 tests across 339 files (1369.46 s); verify exited 0 in 1384 s. Full log: scratchpad/temp/task217/full-e986e8d.log, SHA-256 7efeb51a58f68d88f7b7177671f284165ff3244dc8328704cd6174905e2fe7ad.
• Independent read-only review found no remaining material blocker after unique invocation-directory and exact numeric-schema corrections.

AC1's exact historical substage, the fix decision and AC3's ten hosted main passes per OS remain open. This PR collects future evidence; it does not claim the wall misses are resolved.

Retirement condition: remove the reported-only transcript phase diagnostic, its exact-source phase anchors and unit controls, and its CI step when #367 closes. Ordinary product refactors must update these anchors while the diagnostic remains in use. The repaired SessionStart probe keeps its fail-closed drift guard.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-11T03:15:07.465552Z	e986e8d	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.