docs(adr): record the ruled ADR-0076 D11 revision — the rest-server half is delineated, not split

[os-project-manager]

Fixes #9024

The maintainer ruled option B on #5949 (2026-08-15, verbatim 「接受你的建议」, recorded at comment 5301726968): the ADR-0076 D11 second half — packages/rest/src/rest-server.ts — is delineated, not split. B's own deliverable was an ADR-0076 D11 revision saying so, and it was never written, so D11 still read as an unmet obligation: a reader arriving at the ADR concluded there was undone work, when in fact there is a decision that it will not be done.

This writes the decision down. No code moves; not one line of rest-server.ts. ⛔ This does not re-open A/B.

What changed in docs/adr/0076-objectql-core-tiering.md

1. Header status line — D10/D11/D12 partially landed becomes D10/D12 partially landed plus an explicit D11 clause: implemented on the dispatcher half, decided on the rest-server half.
2. The D11 status bullet — "the rest-server half is untouched" becomes "decided: delineated, not split", pointing at the revision. Two deliberate calls here:
   • the bullet's raw-LOC comparison ("now ~7.7k LOC, i.e. larger than the ~5.1k figure below") is removed, not refreshed — the ruling retired that proxy, so refreshing it would re-canonise the retired instrument in the very line that records its retirement;
   • the dispatcher figure in the same bullet was stale and is re-measured: http-dispatcher.ts is 2,087 lines at dc280d326, against the ~1.7k the 2026-08-05 calibration recorded. In-scope because it lives in the sentence being rewritten; called out here because it is a figure change nobody asked for.
3. A new revision block at the end of D11 carrying the four things that make the decision reviewable rather than merely asserted — the structural reading, the accepted price, the replacement metric with its reopen condition, and the one genuinely undecomposed mass.
4. The withdrawn "second central route generator" wording, which is inside D11's own decision text, is marked in place as withdrawn. Left rather than deleted because it is what the decision was written against; marked because a revision resting on the opposite structural reading is unreviewable while the contradicting sentence stands unannotated.

The accepted price is stated, not implied

B is defensible and it is not free. Merge conflicts, affected-subset CI, and the PM lane's batch-independence rule all operate on file paths, so a delineation expressed as methods inside one file is invisible to all three: two agents editing registerApprovalsEndpoints and registerReportsEndpoints — blocks that share no line — still collide. The revision says that plainly, and says it is permanent until someone reverses the decision. An ADR recording a decision without its cost is one a future reader re-opens.

Three withdrawn claims are marked withdrawn, never inherited

An ADR revision is exactly where a retracted claim gets accidentally canonised.

• "the growth curve shows the interim convention is not holding" — not carried, in any form, including as background. The revision carries only the composition measurement that retires the series.
• "this file costs the lane serialisation roughly once a round" — recorded as the impression it was, and withdrawn; the measured number is reported as the measured number.
• "a central route generator" / "a route table to split" — not reused; marked withdrawn where D11 still carries it.

Every figure re-derived against origin/main @ dc280d326

Nothing was copied from the card. Method: git diff/wc plus a comment-aware line classifier, and the same classifier was re-run on the tree the ruling measured, so the two readings are like-for-like.

Figure	Card / ruling	Measured at dc280d326	Verdict
registerXxxEndpoints methods	18	18	holds
split on the same axis as packages/runtime/src/domains/*	asserted	holds — 15 domain modules behind DomainHandlerRegistry.list() there; 18 methods behind one 73-line registerRoutes() here	holds
whole-file composition	6,008 code / 5,400 comment-blank (11,408 total)	5,665 code / 4,690 comment-blank (45.3%), 10,355 total	superseded
registerMetadataEndpointsInner	~2,328 lines	2,545 lines	superseded
replacement metric, first datapoint	0	0, and 0 again for the 2026-08-16 round	holds

Why the composition figure moved, stated rather than reconciled: the ruling's numbers were taken at 5047cb8ab, before #8850's prologue extraction (landed as 8664a2c99) moved 1,296 lines of ADR-0112 error/fault classification out to packages/rest/src/error-response.ts. The same classifier reproduces 6,008 / 5,400 exactly at 5047cb8ab — so the difference is the file changing, not the method. registerMetadataEndpointsInner reads 2,324 on that tree against the quoted ~2,328 (a method-boundary convention, not a disagreement) and 2,545 today. Both are written as measured; neither is averaged or quietly adjusted.

How the metric's datapoint was re-derived, since it is the one figure not readable from the tree: #8087 — the one queued card that would land in rest-server.ts — is still open and still pm:blocking for a reason unrelated to this file, so the 2026-08-15 collision never materialised. For the current round, none of the 24 open pm:dispatched cards lands in rest-server.ts (the only packages/rest-adjacent one, #8739, is packages/types/src/error-leak.ts). Both rounds read 0.

Verification

Gate union re-derived from the actual changed path after the final commit and run at d53b0f5ed (= the pushed head):

node scripts/pm/dispatch-gates.mjs docs/adr/0076-objectql-core-tiering.md
  -> pnpm check:adr-anchors · node scripts/check-adr-links.mjs · node scripts/check-adr-merge-approval.mjs

check-adr-anchors: OK (49 anchored file(s); 121 decision number(s); 25777 citation(s) across 3212 file(s) resolve)
check-adr-links:   OK (541 relative link destination(s) under docs/adr/ resolve)
check-nul-bytes:   OK (scanned 5956 text file(s); no raw ASCII control bytes)

check-adr-merge-approval.mjs is the ADR maintainer approval gate. It cannot pass from a dev seat by design — it reads PR reviews and needs a token this seat does not carry (401 locally). ⛔ No attempt was made to satisfy it by relabelling; it is CI's and the maintainer's.

No changeset: docs/adr/** publishes nothing, so skip-changeset applies.

⛔ Merge posture

ADR-class path. Draft, unqueued, no auto-merge, per the standing rule that AI seats do not merge or queue ADR PRs (#6741 / #7548 / #7623). It waits for a human.

---

Generated by Claude Code

[os-project-manager]

CI note from the dispatching PM seat: the red ADR maintainer approval is expected and will not be "fixed" — it is the gate doing its job.

scripts/check-adr-merge-approval.mjs reads PR reviews and requires a human maintainer's approval on any docs/adr/** path. No AI seat can turn it green, and none should try: this repo's standing rule is that AI seats do not merge, queue, or auto-merge ADR-class PRs (#6741 / #7548 / #7623). ⛔ I have not armed the merge queue on this PR and will not. It waits for a human.

The rest of the derived union is green at the pushed head d53b0f5ed: check-adr-anchors OK (49 anchored files, 121 decision numbers, 25,777 citations across 3,212 files resolve), check-adr-links OK (541 relative destinations under docs/adr/ resolve), check-nul-bytes OK.

Two things a reviewer should look at first, because both are figure changes beyond the literal ask and both are declared rather than buried:

1. The whole-file composition figure moved — the ruling's 6,008 code / 5,400 comment-blank reads 5,665 / 4,690 today. The PR does not reconcile the two: it re-runs the same classifier at the ruling's own tree (5047cb8ab) and reproduces 6,008 / 5,400 exactly, then attributes the delta to Extract rest-server.ts's error/fault-classification prologue — it is ADR-0112's concern, not the REST server's #8850's prologue extraction moving 1,296 lines out to error-response.ts. That is the difference being the file changing, not the method disagreeing — and it is demonstrated rather than asserted, which is the standard this class of edit needs.
2. The dispatcher figure was stale and is re-measured (~1.7k → 2,087 lines). Nobody asked for it; it lives inside the sentence being rewritten, and the PR body says so plainly instead of letting a reviewer discover an unannounced number.

Also worth a reviewer's eye: the withdrawn "second central route generator" wording is marked in place rather than deleted, on the reasoning that a revision resting on the opposite structural reading is unreviewable while the contradicting sentence stands unannotated. That is a judgment call about ADR history and is exactly the kind of thing the human merge gate exists to arbitrate.

---

Generated by Claude Code