ADR-0076 D11 needs its ruled revision: the rest-server half is delineated-not-split, and the ADR still records it as an unmet obligation

[qq9340100]

Filed unassigned and unlabelled for triage routing by the domain:cli PM seat (session session_01NaS1PAHJcPfAA2acnV53Tn). Not a claim. This lands in docs/adr/**, which is ADR-class and a path this lane's dispatches fence off — flagging rather than silently taking it.

What is owed

The maintainer ruled option B on #5949 — the ADR-0076 D11 second half (packages/rest/src/rest-server.ts) is delineated, not split; the decomposition programme does not open.

• Provenance: maintainer, live PM chat 2026-08-15, verbatim 「接受你的建议」, accepting the integrated recommendation in #5949 comment 5301723574. Ruling recorded at comment 5301726968.

B's own deliverable, as the 2026-08-08 ruling framed it, is "an ADR-0076 D11 revision declaring the second half delineated-not-split". That revision has never been filed as a card and has never been written. Searched open and closed issues before filing; the near neighbours are #5357 (D11 anchors, closed), #5949 (the finding itself, still open), and #8850 (the prologue extraction, closed). No duplicate.

Why this is not bookkeeping

Right now docs/adr/0076-objectql-core-tiering.md D11 still reads as an unmet obligation — its status line says the rest-server half is untouched and larger than the figure it records. A reader arriving at that ADR today concludes there is undone work. There is not: there is a decision that it will not be done, recorded only in an issue comment thread.

That is the same defect #5949 was filed about in the first place. Its opening line is 「缺的是承接」 — an observation with no written home. Ruling B without revising the ADR reproduces it exactly one layer up: the decision now has no承接 either, and the next ADR calibration round will rediscover "D11's second half never landed" and re-litigate a settled question. This card exists so that does not happen.

Scope

Revise ADR-0076 D11 so its status changes from not implemented to decided: delineated, not split. The revision should carry, because these are what make the decision reviewable rather than merely asserted:

• The structural reading the ruling actually rests on — RestServer registers routes through 18 registerXxxEndpoints methods that are already split by domain, i.e. the same axis packages/runtime/src/domains/* uses. The file is not an undifferentiated blob, which is why "already delineated" is an honest description rather than a rationalisation.
• The accepted price, stated plainly — the domain delineation stays invisible to merge conflicts, affected-subset CI, and lane serialisation, because all three operate on file paths. Two agents editing registerApprovalsEndpoints and registerReportsEndpoints — blocks sharing no line — still collide. B is defensible; it is not free, and the ADR should say so rather than let a future reader discover it.
• The replacement metric and the reopen condition — the line-count series is retired (it counted a proxy that is ~47% prose). It is replaced by: per PM round, the number of times two dispatchable cards could not run concurrently because both land in rest-server.ts. Reopen A if that counter sustains above ~1/round. First measured datapoint: 0.
• The one genuinely undecomposed mass, so B is not read as blessing everything in the file: registerMetadataEndpointsInner at ~2,328 lines with no internal delineation. Its characterization is Characterize registerMetadataEndpointsInner (2,328 lines, 20% of rest-server.ts) — read and propose, do not split blind #8851's subject and is not foreclosed by this revision.

⚠️ Two corrections that must not be inherited

Both are errors the seat made and then measured its way out of on #5949. They are recorded here because an ADR revision is exactly where a withdrawn claim gets accidentally canonised:

1. "The growth curve shows the interim convention is not holding" — withdrawn. It used raw wc -l, and the file measures 6,008 code lines against 5,400 comment/blank (~47% prose), with the comments being dense explanatory documentation rather than commented-out code. A raw line delta cannot separate "new endpoints landed inline" from "someone documented a subtle refusal properly". ⛔ Do not carry the line-count series into the revision as evidence.
2. "This file costs the lane serialisation roughly once a round" — an impression, not a measurement, and the first real datapoint contradicted it (0). ⛔ Do not restate it. The zero strengthens B rather than weakening it, but it has to be reported as the measured number it is.

Also do not carry the old framing that D11's second half is "a central route generator" / "a route table to split" — measured false, and it survived five triage rounds before anyone checked.

Not in scope

⛔ No code moves. ⛔ The registerMetadataEndpointsInner characterization is #8851, not this. ⛔ This card does not re-open A/B — the ruling is made; this writes it down.

Related: #5949 (the finding and the full four-prism analysis) · #8850 (prologue extraction, landed) · #8851 (characterize registerMetadataEndpointsInner) · #5357 / PR #5940 (D11 anchors) · #5063 (D1–D12 status-line calibration) · ADR-0076 D11 / OQ#9 · ADR-0112 (the prologue's real owner).

[os-project-manager]

Triage (triage seat, session session_01VUKiXCWoLo8Vnd26tHqXzj): lands in docs/adr/0076-objectql-core-tiering.md ⇒ domain:devx, pm:queue, type Task.

Rationale: writes down an already-made ruling (Option B on #5949, maintainer 2026-08-15, verbatim 「接受你的建议」, recorded at #5949 comment 5301726968) — no re-adjudication, so this is dispatchable. The body already carries the full revision content, the two withdrawn claims that must NOT be inherited, and the reopen metric.

⚠️ ADR-class path (docs/adr/**): the dev may draft, push, and open a draft PR, but per standing rule (maintainer 2026-08-08) the PR must be left for the maintainer to merge — reviewing seat applies the accept-with-human-merge path (no ready-flip into the queue, no auto-merge) and lists it as awaiting a human merge.

Size/model suggestion: S–M, single-file ADR prose with the content pre-specified; opus.

---

Generated by Claude Code

[os-project-manager]

Claim (dev seat, session session_011RB4waLuNbdruCo6X9oobm), branch claude/issue-9024-adr0076-d11-revision, worktree ../objectstack-9024 cut from current origin/main.

Re-read the comments before starting: only the triage comment (session session_01VUKiXCWoLo8Vnd26tHqXzj) is present — no earlier dev claim, so this is unclaimed and I am taking it.

Plan: single-file revision of docs/adr/0076-objectql-core-tiering.md D11, no code moves. Every figure the card carries gets re-derived against origin/main before it enters the ADR (the 18 registerXxxEndpoints methods and their axis, the code-vs-comment split, registerMetadataEndpointsInner's size, and the replacement metric's first datapoint); a figure that has moved will be written as measured with the card's value called stale, not silently reconciled. Draft PR, left unqueued for a human merge per the ADR-class rule.

---

Generated by Claude Code

[os-project-manager]

{
  "issue": 9024,
  "status": "done",
  "branch": "claude/issue-9024-adr0076-d11-revision",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/9057",
  "premise_still_valid": true,
  "summary": "ADR-0076 D11 now records the 2026-08-15 ruling instead of an unmet obligation. Three edits to docs/adr/0076-objectql-core-tiering.md, no code moved: the header status line and the D11 status bullet change from 'the rest-server half is untouched' to 'decided: delineated, not split', and a revision block at the end of D11 carries the structural reading (18 registerXxxEndpoints methods enumerated by one 73-line registerRoutes(), the same axis packages/runtime/src/domains/* splits on), the accepted price stated plainly and marked permanent (merge conflicts, affected-subset CI and lane serialisation all key on file paths, so registerApprovalsEndpoints vs registerReportsEndpoints still collide), the replacement metric with its reopen condition and its measured 0, and registerMetadataEndpointsInner named as the one undecomposed mass that #8851 owns and B does not bless. All three withdrawn claims are marked withdrawn rather than inherited, including the 'central route generator' wording that D11's own decision text still carries, which is annotated in place. Two deliberate calls worth the PM's eye: the status bullet's raw-LOC comparison was REMOVED rather than refreshed (refreshing the retired proxy in the line recording its retirement would re-canonise it), and the same bullet's stale dispatcher figure was re-measured to ~2.1k (2,087 lines) from the ~1.7k the 2026-08-05 calibration recorded — in scope because it sits in the sentence being rewritten, but it is a figure change nobody asked for.",
  "tests": "No code change, so no package test/typecheck applies. Gate union re-derived from the actual changed path AFTER the final commit and run at d53b0f5ed (= pushed head, git rev-parse --short HEAD): `node scripts/pm/dispatch-gates.mjs docs/adr/0076-objectql-core-tiering.md` -> check:adr-anchors, check-adr-links.mjs, check-adr-merge-approval.mjs; plus check:nul-bytes (any edit). Results: 'check-adr-anchors: OK (49 anchored file(s), every governing ADR still referenced; 121 decision number(s); 25777 citation(s) across 3212 file(s) resolve)'; 'check-adr-links: 541 relative link destination(s) under docs/adr/ resolve'; 'check-nul-bytes: OK (scanned 5956 text file(s) -- 5956 tracked, 0 untracked-not-ignored; skipped 5 binary; no raw ASCII control bytes)' — each self-test run first. check-adr-merge-approval.mjs fails locally by design (HTTP 401 Bad credentials reading PR reviews; it is the ADR maintainer-approval gate) — no attempt made to satisfy it. Control-byte self-scan beyond the gate: `grep -naP '[\\x00-\\x08\\x0b\\x0c\\x0e-\\x1f\\x7f]' docs/adr/0076-objectql-core-tiering.md` -> no matches. FIGURE RE-DERIVATION, all against origin/main @ dc280d326 with a comment-aware classifier (nothing copied from the card): 18 registerXxxEndpoints methods HOLDS (still exactly 18; 17 domain entry points plus registerMetadataEndpointsInner, which registerMetadataEndpoints wraps for its audience gate). Same axis as packages/runtime/src/domains/* HOLDS (15 domain modules there behind DomainHandlerRegistry.list(), plus a shared non-domain unavailable.ts; 18 methods here behind one 73-line registerRoutes()). Composition 6,008 code / 5,400 comment-blank is STALE — measured 5,665 code / 4,690 comment-blank (45.3%), 10,355 total; the same classifier reproduces 6,008/5,400 EXACTLY at 5047cb8ab, the tree the ruling measured, so the readings are like-for-like and the delta is #8850's prologue extraction (landed as 8664a2c99) moving 1,296 lines out to packages/rest/src/error-response.ts. registerMetadataEndpointsInner ~2,328 is STALE — 2,324 on that same tree (method-boundary convention, not a disagreement) and 2,545 today. Metric first datapoint 0 HOLDS, re-verified not restated: #8087, the one queued card that would land in this file, is still open and still pm:blocking for an unrelated reason, so the collision never materialised; re-derived independently for the 2026-08-16 round, also 0 — 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 stale figures are written as measured and flagged superseded in the ADR and PR body; neither is averaged or quietly reconciled. skip-changeset applied by reading the PR's labels then writing the union; read back after the write: [\"size/s\",\"skip-changeset\"], still draft:true.",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #9058: ADR-0076 OQ#9's D11 worklist records outcomes for items (1)(2)(3) but never for (4) — the KernelResolver env-resolution seam exists and owns env resolution in the dispatcher (attributed to ADR-0006 Phase 5 in source), while ADR-0076, which asked for it, records nothing; observation class, `finding` label, no pm:queue, unassigned"
  ]
}

Merge posture (unchanged by me): PR #9057 is a draft, unqueued, no auto-merge — ADR-class path, waits for a human per #6741 / #7548 / #7623.

Spot-checks that came back clean, so nothing was filed for them: D1's claim that SysMetadataEngine never reached packages/metadata-core/src still holds (zero occurrences), and D10's claim that ObjectStackProtocolImplementation still exists holds (packages/metadata-protocol/src/protocol.ts:3494).

---

Generated by Claude Code