packages/spec/src/type-alias-convention.pin.test.ts serialises the spec lane — its dense sequential `M…` / `Iso…` numbering makes any two concurrent schema-adding PRs conflict

[os-warren]

Filed by domain:spec execution seat 2, session session_01UDXER3sdqfeVYpEWZs5mZx, 2026-09-22T05:28Z. ⛔ Unlabelled and unrouted on purpose — an execution seat files, triage grades and routes.

This is not a product defect and does not claim to be one. It is a lane-throughput defect with a named landing point, and it has already cost this lane a concrete thing today, which is why it is a card rather than a note.

The collision that happened, bisected rather than asserted

All readings taken 2026-09-22T05:2xZ. Instrument: git merge-tree --write-tree --name-only BASE HEAD (read-only, writes no working tree), against a freshly fetched origin/main @ 8f30c188a364a9cc982059f803914ee9d4826615.

PR #19600 (card #15178, domain:spec, size/xl, parked on a maintainer authorization) head 8dcd6a42ae323931b02e12208bf229da921b9eb3:

CONFLICT (content): Merge conflict in packages/spec/src/type-alias-convention.pin.test.ts

One file, and only that file. Bisected against main:

base	merge-tree	verdict
1b82c519df (#19226)	rc=0	clean
eff0a9622340f59068651f2b82d4afcd177546a8 — parent of the next row	rc=0	clean
5ce3705052 — PR #19638, merged 2026-09-22T04:52:12Z	rc=1	CONFLICT

⇒ the conflicting commit is identified, not guessed. And it is this seat's own landing. #19638 is a PR this seat reviewed and enqueued earlier in the same session, on card #17551, and the miss is in this seat's own claim 5769968983 — read back first-hand for this card, not recalled:

File surface: packages/spec/src/contracts/analytics-service.ts (or packages/spec/src/api/analytics.zod.ts — the ruling leaves the placement to this seat, one place), the POST /api/v1/analytics/dataset/query door, their tests, regenerated artefacts, .changeset/.

Two defects there, both this seat's:

1. That claim carries no Serial constraints cleared: line at all. Sibling claims in this lane on the same day do (e.g. domain:spec#4 on 14 ADR-0049 tombstones tell authors a key was removed in @objectstack/spec 18 — an npm version that does not exist and, under the 2026-09-13 level ruling, will not be the carrier #18048 and on A multi-package artifact serializes its metadata twice — the flattened top level and every packages[i] body carry the same definitions #14512). So no serial census stood between feat(spec): author DatasetSelectionSchema and parse the whole selection at the analytics dataset door #19638 and feat(spec)!: split the translation bundle type — settings is a platform group, not a per-app one (#15178) #19600's open surface — ⛔ this is not a census that ran and missed, it is a census that was not written.
2. 「regenerated artefacts」 was doing work it cannot do. type-alias-convention.pin.test.ts is 「Generated once … maintained by hand from here on」, so a dev could in good faith read it as covered by that phrase and edit it without the edit ever surfacing as a widening. ⛔ That is the declaration's vagueness, not the dev's error.

It is stated here rather than only on the card because a reader deciding this card's worth should know the collision came from two PRs in the same lane, and that tightening the claim template would not have prevented it — the file is one every spec PR must touch, so the serial queue for it is the whole lane.

Lit control for the instrument (so the rc=1 is not an instrument that only ever says CONFLICT): PR #19643 head c9058f4056193ddeb7535924ba53d577b12d4a2b, same instrument, same base, same act ⇒ rc=0, clean. PR #19598 head cfc103a20a34a6eb35670986a130fa7b3dc266ff ⇒ rc=0, clean.

Why it is structural, in the file's own words

packages/spec/src/type-alias-convention.pin.test.ts is 2349 lines on origin/main @ 8f30c188a3, and its header says what it is:

Generated once from the measured corpus (see ADR-0122's appendix); maintained
by hand from here on — a line leaves when its schema gains a shape, and the
gate refuses a bare alias that is neither pinned here nor paired with an
XParsed.

It is also the registry scripts/check-spec-parsed-alias.mjs reads, so it is not optional and not skippable: every PR that adds a schema must edit it.

The two sides of today's conflict touch the same regions, which is the whole point:

region	main side (#19638)	PR #19600 side
import block	@@ -275,7	@@ -275,7
pin block	@@ -345,6	@@ -1251,6
describe body	@@ -1683,7	@@ -1683,7
describe body	@@ -2268,7	@@ -2268,7

Three of the four regions are byte-identical starts. This is not two PRs unluckily landing near each other — it is two PRs doing the one thing this file requires of everyone.

What makes it collide, said as the thing to delete

The entries are keyed by a dense sequential counter, in two independent series:

• modules: import type * as M187 from './shared/duration.zod.js';
• pins: export type Iso651 = Assert<Eq< z.input< typeof M152.… >, z.infer< … > >>;

Two PRs that each add one schema each allocate the next number in both series, at the end of both blocks. Git has nothing to merge on: same line, two different contents. The counter is the construct that permits the error.

Following 「先删容许出错的构造,再让正确形态成唯一拼写,最后才加检查」, in that order:

1. Delete the counter. Key both series on the thing that is already unique and already stable — the module path and the schema name (M_sharedDuration, Iso_ActionResultDialogTranslation). Two PRs adding different schemas then write different lines.
2. Make the correct spelling the only one. Keep both blocks sorted by that key, so the insertion point is a function of the name rather than of arrival order, and two concurrent inserts land in different places and auto-merge.
3. Only then consider a check that the blocks are sorted.

⛔ Not proposed, and deliberately so: regenerating the file wholesale, weakening any assertion, changing what scripts/check-spec-parsed-alias.mjs accepts, or changing the exemption set. The assertion set before and after must be identical member-for-member, and that identity is the acceptance condition. A renumbering PR that silently drops one pin is worse than the collisions it fixes.

Materiality, with its limit stated

Commits on origin/main touching this file: 3, dated 2026-09-18, 2026-09-20, 2026-09-22.

⚠️ That is a lower bound, not a rate. This checkout is shallow (git rev-parse --is-shallow-repository ⇒ true; 303 commits reachable; oldest reachable commit dated 2026-09-18), so the window is 4 days and the count cannot see past it. Whoever grades this card should re-take the count on an unshallow clone before treating 3 as the number.

What does not depend on that count: the collision is demonstrated, and its cost is currently being paid — PR #19600 carries an at-tier contract-review PASS pinned to head 8dcd6a42ae, and clearing this conflict moves that head and strands the record.

Duplicate-search words

type-alias-convention.pin.test, check-spec-parsed-alias, ADR-0122, sequential alias counter, serial collision

---

Generated by Claude Code

[os-warren]

⚠️ Two corrections to this card from its own resolution round — one narrows it, the other makes it WORSE than filed

Filed by domain:spec execution seat 2, session session_01UDXER3sdqfeVYpEWZs5mZx, 2026-09-22T06:48Z. Both corrections are measured, both contradict the card as written, and both come from the round that resolved the collision the card was filed about (PR #19600, head 8dcd6a42ae → 34b6a25b2c).

Correction 1 — ⛔ only ONE of the two series collided. The card claimed both.

The card says: 「The entries are keyed by a dense sequential counter, in two independent series」 — modules M187 and pins Iso651 — and its remedy prescribes re-keying both.

Measured. Module-path sets, across four trees:

tree	module paths	pins
merge-base 1ff3a8f210	174	784
PR head 8dcd6a42ae	174	786
origin/main e9861d217a	174	787
merged 34b6a25b2c	174	789

Zero difference in the module set, anywhere — because both sides added their schemas to modules that were already imported (M152 on the branch, M11 on main). ⇒ only the Iso counter collided, at Iso877 and Iso878.

⭐ And this seat had the reading in hand and did not draw the inference: its own independent union measurement printed modules=174 on every tree and it still relayed the two-series framing. The dev drew it. ⇒ the module half of this card is a conjecture, not a demonstration — it would take two PRs each importing a new module, which has not been observed. ⛔ Whoever takes this card should not present the module series as a measured collision, and should decide on its own evidence whether re-keying it is worth doing at all.

⭐ Correction 2 — the collision carried NO conflict marker, and that is worse than what the card describes

The card says: 「Git has nothing to merge on: same line, two different contents.」 That describes a conflict — something that stops you.

Measured, in the resolution round: the two allocation sites sit roughly 910 lines apart, so git text-merged them cleanly. Git reported three marker hunks in this file, and all three were count prose and narrative — ⛔ none was the identifier collision. The clean merge produced a file declaring export type Iso877 twice and export type Iso878 twice.

⇒ resolving only the marked hunks and committing would have shipped duplicate identifiers. The danger is not that a conflict must be resolved by hand; it is that the collision does not present as a conflict at all. A hand-resolution that looks complete, and a review that reads the marked hunks, both pass it.

What caught it was the round's acceptance condition — the union proof, keyed on module path + schema name rather than on the counters — plus the fact that the pin file is type-checked: packages/spec/tsconfig.json excludes **/*.test.ts, so the package's bare tsc --noEmit does not compile it; coverage comes from tsconfig.test.json via check:test-typecheck, and the file has zero entries in test-typecheck-debt.json. The dev proved that non-vacuous by ablation from the committed state — renaming Iso880 back to Iso877 turned check:test-typecheck red with 「2 type error(s) in a file the ledger does not cover」, restored byte-identically.

⇒ the card's priority should rise, not fall. A construct whose failure mode is a silent clean merge into duplicate identifiers is worth more than one whose failure mode is a conflict, because only the first can reach a reviewer looking correct.

What stands unchanged

The demonstrated collision, its bisection to 5ce3705052 (#19638), the lit controls, the file's own 「maintained by hand」 header, its role as the registry scripts/check-spec-parsed-alias.mjs reads, and the ordered remedy 「先删容许出错的构造,再让正确形态成唯一拼写,最后才加检查」 — all unchanged. So is the ⛔ list: no wholesale regeneration, no weakened assertion, no change to what the gate accepts, and set-equality member-for-member as the acceptance condition.

⚠️ And the seat's self-implicating half stands too: the claim behind #19638 carried no Serial constraints cleared: line, and its 「regenerated artefacts」 phrase covered a file that is maintained by hand.

---

Generated by Claude Code