p0: dispatch-gates' `.changeset/*.md` population control makes `Lint & Repo Gates` red on the Version Packages PR — the release cannot be merged, and main goes red the moment it lands

[hotlong]

What is broken

scripts/pm/dispatch-gates.mjs:13052 asserts, in its self-test, against the live working tree:

const changesetPop = suffixCorpus.filter((f) => /^\.changeset\/[^/]+\.md$/.test(f));
t('the tree really does hold a large `.changeset/*.md` population', changesetPop.length >= 100);

The block's own comment states the intent: "Read from the REAL corpus: a fixture cannot show that the tree still holds the population the trap needs, and this one grows with every merged PR."

A changesets version pass consumes every .changeset/*.md. On the Version Packages PR (#11336) that is 861 files, all removed. So changesetPop.length is 0 there and the control fails:

✗ the tree really does hold a large `.changeset/*.md` population
✗ dispatch-gates self-test: 1 of 1353 case(s) failed.
##[error]Process completed with exit code 1

Measured on #11336 head b8573e843, job 100940300762: 30 success, 1 failure — Lint & Repo Gates.

Why this is p0

1. Lint & Repo Gates is one of the six required contexts (AGENTS.md §7). A required red means the Version Packages PR cannot be merged.
2. Merging that PR is the release trigger (ADR-0125 D1). So this gate currently makes a GA release impossible — not slow, impossible.
3. It also turns main red the moment the release lands. main carries 865 changesets today, so the control passes; a version pass consumes all of them, and the merge puts a tree with ~0 changesets on main. The same assertion then fails on main for every PR until 100 fresh changesets accumulate — shared damage, the class AGENTS.md says to stanch immediately.
4. It is new: the assertion landed 2026-08-31 (eb717a12a, fix(pm): follow a final-segment basename glob in hintCovers, and stop the residue naming a move that never happened #13783); the last release was 2026-08-23. This is the first release cycle in which it exists, which is why nobody has hit it.

The shape of the fix (not a ruling — measure it)

The control exists to stop the two live-specimen assertions below it passing vacuously on an empty set, which is a real thing to guard. What it cannot do is require the tree to hold a population, because the population is zero by design twice in every release cycle (on the version PR, and on main right after it lands) and small for days afterwards.

So the honest shape is probably: when the population is empty, report NOT MEASURED loudly, naming why (this tree has no .changeset/*.md — that is what a version pass produces), and skip only the assertions that need the corpus; keep every population-independent assertion in that block running (the extension pin, the single-* segment pin, the directory-surface pin, the unrelated-root pin, and the four globCarriesLiteralSuffix pins all take literal inputs). A silent skip is the wrong answer — it would let the specimen rot unnoticed, which is the failure the ≥100 was reaching for.

⚠️ Note the threshold has a middle range too: between 1 and 99 changesets the assertion still fails, which is the state main is in for days after every release. Whatever replaces it must be correct at 0, at 3 and at 865.

Found while diagnosing the red on #11336 at the maintainer's request, 2026-09-04, during release preparation.

Generated by Claude Code

[hotlong]

Claim: release-blocking gate fix (maintainer-instructed)
Session: session_62849a16-0144-4728-8942-9f60bb3a73f1
Branch: claude/issue-15255-changeset-population-control
Worktree: objectstack-issue-15255
Domain: domain:devx (scripts/pm/**, objectstack-side single writer)
File surface: scripts/pm/dispatch-gates.mjs (the .changeset/*.md live-specimen block and its self-test) — stop on breach; explain in the report
Container & model: S/M, model: opus
Clause-②: no (a gate self-test; no contract surface, no public API)
Serial constraints cleared: no open PR touches scripts/pm/dispatch-gates.mjs (REST pulls list 2026-09-04); the Version Packages PR #11336 is BOT-generated and regenerated 6-hourly — ⛔ the fix lands on main, never by pushing to changeset-release/main

Maintainer direct instruction in this session, verbatim: 「pr 红了，你帮我解决 https://github.com/objectstack-ai/objectstack/pull/11336」 — direct-dispatch channel (2026-08-10 standing authorization). Diagnosing and fixing the gate is ordinary work; ⛔ the Version Packages PR itself is never merged, queued or auto-merged by any agent seat (Prime Directive #15).

Generated by Claude Code

[hotlong]

Claim: os-dev seat, starting implementation.

Session: session_62849a16-0144-4728-8942-9f60bb3a73f1
Branch: claude/issue-15255-changeset-population-control (pushed empty at start as the write-route probe; push works from this seat)
Worktree: objectstack-issue-15255, cut from origin/main at a23603ea916c2f5065660006119efee0b889e473
File surface: scripts/pm/dispatch-gates.mjs — the .changeset/*.md live-specimen block and its self-test

Fix lands on main as an ordinary PR. This seat never pushes to changeset-release/main, and never merges, queues or auto-merges the Version Packages PR.

Generated by Claude Code

[hotlong]

os-dev-report

{
  "issue": 15255,
  "status": "done",
  "branch": "claude/issue-15255-changeset-population-control",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/15277",
  "premise_still_valid": true,
  "summary": "The failure is real and was reproduced byte-for-byte locally, but one detail of the card's premise is wrong and it changed the fix. The population on the Version Packages PR is 1, not 0: README.md survives a version pass, so b8573e843 carries README.md + config.json (read via the contents API on that ref). That is why CI reported 1 of 1353 rather than 3 failures, and it means a guard written as 'skip when the population is 0' would still have been red on the very tree it was written for. So the size claim was removed rather than gated: the three corpus assertions now run whenever the population is non-empty (real at 1, 3 and 865), an empty population reports NOT MEASURED out loud with its own line, a checkable reason and a tally repeated on the verdict line, and what the gate requires instead is the specimen's HOME (.changeset/config.json tracked) - which separates 'resting mid-cycle' from 'rotted' and is the property that stops a silent skip letting the specimen decay. A SECOND live-population site was found and guarded the same way (the residue block's hintReachesTree); unguarded it would have kept producing the release-blocking red after the measured site stopped. Ten new pins cover the new decision at 0/1/3/865, the home discriminator both ways, and the tally renderer. ANOMALY for the PM, unattributed: partway through the run, 865 .changeset/*.md files vanished from this worktree's DISK (index intact); the self-test was the obvious suspect and was EXONERATED by direct measurement (869 on disk before and after a full run, tree clean). No reading was affected - trackedFiles() reads git ls-files, i.e. the index, which was intact throughout. Files restored, tree clean, nothing lost.",
  "tests": "All readings are real runs of `node scripts/pm/dispatch-gates.mjs --self-test`, populations mutated in an ISOLATED scratch worktree (its own index; this branch's tree was never mutated), each mutation proved on disk by counting `git ls-files` before and after. 865 (pre-merge): exit 0, `✓ dispatch-gates self-test: 1363 cases pass.` 869 (post-merge, final head 3c4728cb5): exit 0, `✓ 1363 cases pass.` 3: exit 0, `✓ 1363 cases pass.` with 0 skipped - the corpus assertions RAN, which is the state where `>= 100` was red. 1 (the real release-PR tree), PRE-FIX script: exit 1, `✗ dispatch-gates self-test: 1 of 1353 case(s) failed.` on `✗ the tree really does hold a large .changeset/*.md population` - an exact reproduction of the CI red. 1, FIXED script: exit 0, `✓ 1363 cases pass.` 0: exit 0, `✓ dispatch-gates self-test: 1359 cases pass. ⊘ 2 subject(s) NOT MEASURED on this tree — the .changeset/*.md live-specimen corpus assertions · the residue block's live .changeset/*.md specimen.` The pre-fix leg is the reverse verification: same tree, same command, only the script swapped; restoration proved by `git diff 965e2dc5d --stat` empty. NO vitest suite exists for this script - its suite IS --self-test, which is why check-dispatch-gates.mjs exists as the CI wrapper. Gate family derived on the final commit with `node scripts/pm/dispatch-gates.mjs` (no paths passed), all green with exit codes captured BEFORE any pipe: check:pm-dispatch-gates, check:nul-bytes, check:declared-population-live, check:ratchet-remedy-authority, check:watch-hint-literal, check:parse-guard, check:entry-guard, check:agent-test-spelling, check:cli-command-ids, check:cross-package-test-inputs, check:pm-half-states, check-self-test-wired, check-whole-set-label-write, check-closing-keyword-parity, check-comment-mask-corpus, measure-self-test-floor. ESLint was run over the WHOLE repo, not narrowed: `eslint . --no-inline-config --format json` = 5877 files, 0 errors, 0 warnings. TWO readings that are NOT verdicts on this change: check:bash32-floor fails 8 of 153 here and fails IDENTICALLY on pristine main in a separate checkout (host-bash artifact of this macOS seat); check-closing-keyword-parity and check-comment-mask-corpus first exited on their own stated prerequisite (no node_modules; the former exits 3 and says so) = NOT MEASURED, not a finding - both re-run green after `pnpm install`. Seat declaration, verbatim from every heavy run: `os-verify-lock: VERDICT command-exit 0 · UNLOCKED (declared) · no usable flock on this host, so the shared verify lock was NEVER taken and NOTHING was serialized · ran 140s (2m20s) · declare it in the PR body`.",
  "mcp_calls": "0 — every GitHub read and write on this card went through gh REST (`gh api`); no MCP GitHub tool was called",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #15278: dispatch-gates' `--changed --commands` CONTROL case reds on any branch with no diff against origin/main — i.e. on every fresh worktree's baseline run, before the first commit. Measured on a zero-edit worktree at a23603ea9: `1 of 1353 case(s) failed`; the identical tree after one unrelated commit runs 1363 pass. Invisible in CI (a PR branch always has a diff). Same class as this card, different trigger; fails exemption ② (the right shape is a judgment call, not a mechanical edit), so it was NOT carried in this PR.",
    "NOT filed, reported to the PM instead — no reproduction, so an issue would be a mystery: 865 .changeset/*.md files disappeared from this worktree's DISK mid-run (index intact, restored with `git checkout HEAD -- .changeset/`, tree clean, nothing lost). The self-test was the obvious suspect and is EXONERATED by direct measurement — 869 on disk before AND after a full run, `git status` clean. No reading was affected. This container runs many parallel worktrees; if another seat is running a changesets version pass with a wrong cwd, that is worth knowing before the release.",
    "NOT filed, local-only: check:bash32-floor fails 8 of 153 cases on this macOS seat ('the probe is shell this host can parse'), identically on pristine main in a separate checkout. A property of this host's bash 3.2, green on CI's Linux bash — noise, not a defect."
  ]
}

Generated by Claude Code

[hotlong]

PM attribution for the "865 changesets vanished from disk" anomaly in the os-dev report — it was me, and it is not a release hazard.

The report records it unattributed and flags the scary reading: "if another seat is running a changesets version pass with a wrong cwd, that is worth knowing before the release." That reading is wrong, and leaving it standing would send someone hunting a ghost during release prep.

What actually happened: as reviewer of record I ran a verification leg inside this card's own worktree (objectstack-issue-15255) at ~15:52, to check the fix at an empty population:

mkdir -p /tmp/cs-park && mv .changeset/*.md /tmp/cs-park/
node scripts/pm/dispatch-gates.mjs --self-test
mv /tmp/cs-park/*.md .changeset/

That is the disappearance and the restoration, in the dev's tree, from the PM seat — not a stray version pass, not another seat, and nothing to do with the self-test. git status --porcelain was empty after the restore, which matches the dev's own reading.

Two corrections that follow, both mine:

1. My leg measured nothing. trackedFiles() reads git ls-files — the INDEX — so moving files on disk without touching the index left the population at its full size. The self-test I ran was still seeing 869. I reported that reading as a verification and it was not one; it is withdrawn. The dev's exoneration of the self-test is correct and was reached independently.
2. I re-did it properly afterwards, the way it should have been done in the first place: a temp worktree at the Version Packages PR head b8573e843, then the same command with only the script swapped — main's gate exits 1 on ✗ the tree really does hold a large .changeset/*.md population (an exact reproduction of the CI red), the fixed gate exits 0 with 1363 passing. That is the reading that stands, and it agrees with the dev's.

The dev's premise correction is also the more important one and it lands against my dispatch brief: I briefed "skip when the population is 0", and the population on that tree is 1 (README.md survives a version pass), so the shape I asked for would have shipped still-red on the very tree it was written for. The dev measured it, said so, and built the right thing instead. Recorded here because a brief that was wrong should not disappear into a green PR.

Generated by Claude Code

[hotlong]

ACCEPT — reviewer of record: PM session session_62849a16-0144-4728-8942-9f60bb3a73f1.

Checked against GitHub, not the report: draft PR #15277 targets main, first line Fixes #15255, no other closing keyword, one changed file (scripts/pm/dispatch-gates.mjs, +159/-10) inside the claimed surface, no governed path, skip-changeset present (correct — scripts/pm/** publishes nothing), 34 check-runs green at head 3c4728cb5.

Independently reproduced by the PM before the report arrived, on a temp worktree at the Version Packages PR head b8573e843 with only the script swapped: main's gate exits 1 on ✗ the tree really does hold a large .changeset/*.md population — the CI red byte-for-byte — and this branch's gate exits 0 with 1363 passing. That is the claim this card exists to make, measured on the real tree rather than a fixture.

Three things the dev got right that the dispatch brief did not:

1. the population there is 1, not 0 (README.md survives a version pass), so the brief's "skip at 0" would have shipped still-red;
2. a second live-population site (the residue block's hintReachesTree) would have kept the release-blocking red alive after the first was fixed;
3. the gate now requires the specimen's HOME (.changeset/config.json tracked) instead of a size, which separates "resting mid-cycle" from "rotted" — the property that stops a silent skip letting the specimen decay.

Not verdicts, recorded as read: check:bash32-floor fails identically on pristine main on this macOS seat (host bash 3.2 artifact, green on CI Linux); two gates first exited on their own stated prerequisite and re-ran green after pnpm install. Follow-up #15278 filed for the sibling control case, correctly not carried here.

Flipping to ready and arming queue merge. ⛔ The Version Packages PR #11336 is untouched by this seat, as always.

PR: #15277

Generated by Claude Code