mcporter 0.14.1 cutover record corrected (leftover 0.13.13 daemon retired); socraticode mixed-version procedure - #345

What

This PR is on hold for the v2026.09.26.1 release window, until session 9a sends WINDOW DONE. It edits evidence/receipts/mcporter-0141-qualification-20260925.json, which release_due tracks.

1. mcporter 0.14.1: cutover record corrected, leftover 0.13.13 daemon retired

The 2026-09-25 cutover record said two things that were wrong:

• "No mcporter daemon was running before the switch";
• "the next call started a 0.14.1 daemon".

On 2026-09-26 a read-only check found PID 129979 still running tools/mcporter-0.13.13 … daemon start --foreground on the production socket ~/.mcporter/daemon/user.sock, while bin/mcporter named 0.14.1:

• startedAt was 20:09:44Z on 09-25;
• the socket mtime was the same;
• the daemon's two keep-alive socraticode servers were last used at 20:09:44Z and 20:09:45Z.

The coordinator's session record gives the cause, though the record itself is not retained. My own pre-switch baseline call, mcporter list socraticode, went through 0.13.13 at 20:09:43.395Z, in the same command as the relink. That was after a 20:09:11Z precondition check had found no daemon. So the precondition was stale by the time of the switch.

• Fix, retained in evidence/artifacts/mcporter-0141-daemon-correction-20260926/retire.txt.
  • The run first asserted the PID's 0.13.13 command line and that activeCalls summed to 0.
  • The native mcporter daemon stop through 0.14.1 exited 0. The PID, its two children and the socket were then gone.
  • A production call socraticode.codebase_health exited 0 and started PID 744797 on tools/mcporter-0.14.1, which is now the only mcporter daemon process.
• The receipt gets an addendum, and the decision record gets a dated correction in its mcporter section.
• limitations[4] ("a mixed-version daemon cutover" not exercised) still holds. This was a retirement, not a qualification of mixed-version operation.

2. socraticode 1.15.0: mixed-version source review and cutover procedure

This adds a dated update to the socraticode section of docs/decisions/2026-09-25-workstation-sota-refresh.md. It is a pinned source review at v1.14.0 (2218f251) and v1.15.0 (f6191f07); nothing was run.

• The cross-process coordination code is byte-identical between the two tags: lock.ts, watcher.ts, indexer.ts, graph-inputs.ts, graph-analysis.ts, code-graph.ts, symbol-graph-store.ts and startup.ts.
• Locks and watcher. Locks are proper-lockfile directories under os.tmpdir()/socraticode-locks. They go stale after 120 s, refresh every 30 s and are not retried. There is one watcher per project, whatever its version. Mixing versions causes no lock stealing and no duplicate writes.
• Mixed-version effects:
  • Graph flip-flop: graph-inputs.ts:609 checks the builder's version with a strict inequality, so each version rebuilds a graph the other built.
  • Cleanup race: at startup, cleanup of old symbol-graph generations is coordinated only inside one process (coordinateProject, an in-memory promise chain). A server that starts during another's build may delete that build's staging generation. This comes from reading the code and was not reproduced; codebase_graph_build recovers from it.
  • codebase_stop SIGTERMs the index-lock holder, even when it belongs to another session.
• Decision unchanged. The cutover waits until at least 2026-10-01T11:01Z, the end of the 7-day cooldown. The update records a 6-step coordinated procedure for that date.

Checks

• scripts/validate.py (uv, py3.13) returned {"components": 69, "hashed_files": 6892, "profiles": 4, "receipts": 158, "status": "passed"}. That covers integrity and scope only.
• The component matrix and grand list were regenerated with no change.
• core.hooksPath is scripts/git-hooks, and gitleaks found no leaks.

Review

The Codex cross-family review (codex exec -s read-only, gpt-6-astra, high) of the first head returned DISAGREE with 5 findings:

1. Which call started the daemon, and that its children were socraticode 1.14.0, were stated as fact without retained output.
2. The watcher and write guarantees were overstated.
3. codebase_stop was described as always sending SIGTERM.
4. Lock staleness was confused with lock-directory removal.
5. Because of 4, the procedure waited for the lock directory to disappear.

The fixes are in 2361d77c:

• the two unretained claims are now attributed to the coordinator's records;
• the guarantees are narrowed to what the lock protects;
• the proper-lockfile stale-removal behaviour is cited (v4.1.2 lockfile.js:67-86);
• step 4 now waits on live holders.

The Codex recheck marked all 5 FIXED and found one new wording defect: contention skips rather than queues (indexer.ts:1052-1062). 0c6ddcc7 fixes it. The recheck also matched the sha256 and byte counts of all 5 files against manifests/evidence.json. There is no third review round, because the review loop is bounded.

SOTA sources

• mcporter v0.14.1 release. Daemon protocol 3 in both versions, per the existing cutover.json rollback text.
• proper-lockfile v4.1.2 lib/lockfile.js, the version both installs carry.
• SocratiCode v1.14.0 and v1.15.0 sources. The update links each cited line to its v1.15.0 blob.
• Addenda style: rtk 0.50.0 hook recipe: exclude git show REV:path in every global-option form and git branch #318 and ai-memory 2.4.1: real Codex capture observed after the /hooks trust (receipt addendum) #344.

🤖 Generated with Claude Code