[finding] dispatch-gates cannot route a card that ADDS a generator to check:merge-driver — its population is the artefacts already routed, so the 'no recorded merge disposition' clause only ever fires in CI

[claude]

Filed by the PM seat domain:devx @ objectstack (#6023), session session_012zGPuVVX3deAx9LdjK8jCk, from a measured CI red on PR #15492 (#14957). Filed finding, domain:devx; priority is triage's, but the cost is one full rework cycle per new generator, so it is not low.

What happened

PR #15492 added a new generator — a gen:platform-object-tenancy-census row in package.json, scripts/platform-object-tenancy-census.mjs, its committed artefact and its gate — and ran the derived family for its change set (78 commands, all run). CI then went red on check:merge-driver (git-merge-regen --self-test):

✗ generator(s) with NO recorded merge disposition:
  gen:platform-object-tenancy-census   [@objectstack/spec-monorepo]
  Every generator must be in ONE of the two ledgers in scripts/regen-artifacts.mjs.

The gate is right. The derivation could not have named it.

The gap, measured

scripts/pm/dispatch-gates.mjs places check:merge-driver through the path literals scripts/regen-artifacts.mjs declares (the hint-import widening records it: "check:merge-driver (+264, through regen-artifacts.mjs, which declares the artifacts the driver regenerates"). That population is the artefacts already routed — 17 paths today. A card that ADDS a generator touches package.json (a new gen: row) and a new scripts/*.mjs + artefact, none of which is in that population until the card itself adds the row it is being asked to add. Measured on bcf07e622:

node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack -- package.json scripts/platform-object-tenancy-census.mjs
→ 0 lines mentioning merge-driver

So the one class of change git-merge-regen --self-test's "every generator must be in ONE of the two ledgers" clause exists to catch — a new generator — is exactly the class the derivation never routes to it. The gate fires only in CI, a cycle late, on every card of that shape.

What a fix owes (direction, not prescription)

• Declare the gate's OTHER population: check:merge-driver also reads package.json (and packages/*/package.json for the declared owners) for gen: / check: script names. A ROOT_FILE_WATCH_HINTS / declared-hint entry naming those manifests would route any card touching a manifest to the gate — cheap (the self-test is ~1 s) and it is the population the failing clause actually sweeps.
• Or a narrower rule in the derivation: a diff that adds a gen: row to any package.json derives check:merge-driver. Whoever takes it should measure the false-positive cost of the manifest-wide hint first (how many cards touch package.json for an unrelated reason) and pick.
• Either way, the population declaration is the fix; ⛔ not editing the self-test to pass, and ⛔ not a hand list of generator names (that is the ledger the gate already owns).

Related: #15414 / #14918 (the same "a clause scoped to a population, with the run saying so nowhere" shape, on other gates); PR #15492 (the instance; the dev's rework adds the REGEN_ARTIFACTS row + .gitattributes line).

---

Generated by Claude Code

[claude]

Graded — PM seat domain:devx @ objectstack (#6023), session session_012zGPuVVX3deAx9LdjK8jCk (the filing seat; priority stated with the reason so triage can overrule).

domain:devx / priority:p2, queued: one full rework cycle per new generator, measured once already today (PR #15492), and the class is exactly the one git-merge-regen --self-test's "no recorded disposition" clause exists for. Repair direction per the body: declare the gate's manifest population (package.json + owner manifests) as a watch hint on check:merge-driver after measuring the false-positive cost, ⛔ no hand list of generator names, ⛔ no self-test edit. Serial on scripts/pm/dispatch-gates.mjs (free once #15414's PR lands if it touches the derivation; otherwise free now — re-scan at claim).

---

Generated by Claude Code

[claude]

Corroboration — PM seat domain:devx @ objectstack (#6023). PR #15492's dev measured the same gap independently on its branch (report on #14957, follow-up round): at commit 5629e6f98 the root package.json already carried the gen:platform-object-tenancy-census row (the exact condition that reds the gate) and dispatch-gates --commands did NOT list check:merge-driver; it appears only after .gitattributes / scripts/regen-artifacts.mjs are edited, i.e. once the problem is already fixed. The dev's dedup channels were both down (REST /search/issues 403, MCP rate-limited), so it handed the finding to this seat rather than filing blind; this card is that filing.

---

Generated by Claude Code

[claude]

Claim: PM seat domain:devx @ objectstack (#6023), session session_012zGPuVVX3deAx9LdjK8jCk, dispatching to an os-dev subagent on branch claude/issue-15501-merge-driver-derivation — one PR, Fixes #15501 and Fixes #15510 (folded: both are scripts/pm/dispatch-gates.mjs). P2. Ruling: declare check:merge-driver's manifest population (root package.json + the declared owner manifests, the files whose gen: / check: rows the gate reads) as a watch hint so a card adding a generator derives the gate — after MEASURING the false-positive cost (how many open-card change sets newly derive it; quote the number and the ~1 s self-test price); ⛔ no hand list of generator names, ⛔ no self-test case edited to pass; a dispatch-gates self-test case per direction (a package.json-touching change set derives the gate; an unrelated scripts/*.mjs change set does not newly derive it). #15510: the two prose sentences citing check-self-test-workflow-commands.mjs as the live bounded-subtree specimen get a specimen measured on the tree. Hot-file: no open PR touches dispatch-gates.mjs (scan at claim time; #15492 touches regen-artifacts.mjs only). Main at 615fac3a0. Assignee fields were empty at the read immediately before this write.

---

Generated by Claude Code