test(gates): give both document-count pins the third copy's non-vacuous floor, and census content/docs/** (non-zero, so no expansion)

[os-sam]

Fixes #7901

ⓘ PM 席改动：本行原为 Part of objectui#7901.。按本车道对本卡的裁定，普查非零时的交付物就是「下限 + 普查 + 提案」，
覆盖扩张另立一张卡（objectui#7966）。所以本 PR 关闭 objectui#7901，剩余的覆盖问题由 objectui#7966 承载 ——
让两张卡同时开着描述同一件剩余工作是重复。另：objectui#7901 这种写法不是 GitHub 认的 closing reference 形式，
必须是 #7901；PM 的 clause-② carrier gate 也靠这个关键词把 PR 与卡配对。

Clause-②: no

Two things, in the order the card fixed: the floor lands, the coverage expansion does not. The census came back non-zero on both axes, so per this lane's standing rule the expansion is deferred to its own card and this PR carries the census plus a proposal.

---

1. The non-vacuous floor (landed)

check-links-workflow.test.ts — the third copy of objectui#7448's document-count pin — asserts the extracted prose is longer than 400 characters before asserting it carries no document count. The two twins did not.

Without the floor the count assertion is vacuous by construction: documentCounts('') is [], so a header the extraction has stopped reading — workflow renamed or deleted, comment markers changed, path moved — passes the pin while nothing at all was scanned. It is the same claim the scan-collapse assertions in both files already make about the document walk, applied to the one surface that had none.

Measured in characters before writing, not inferred from line counts:

workflow header	comment lines	characters the pin extracts	ratio to the 400 floor
doc-fence-languages.yml	71	4696	11.7x
doc-component-types.yml	68	4514	11.3x
check-links.yml (third copy, for reference)	61	3673	9.2x

Both twins clear the floor by an order of magnitude, so this is latent, not live. Nothing in the tree changes verdict today.

The extraction is lifted to module scope beside POPULATION_COUNT, unchanged character for character, so the pin and its new control read one definition — the same reason objectui#7914 lifted the pattern a few hours earlier.

Positive control A — in file

Each twin gains the emptiness floor fires on a header this pin has stopped reading. It demonstrates on synthetic YAML that a comment-less workflow extracts to the empty string, that the count half scores that identically to a header carrying no counts, and that the floor is the half which rejects it. A near-empty (two-line) fixture is asserted too, so a header truncated gradually is not waved through by an emptiness-only check.

Positive control B — the real headers mutated, restored by state

Each real workflow header was emptied on disk, the pin file re-run, and the header restored. The mutation was proven on disk by blob hash before the reading was taken, and the restore proven by state — never by an exit code.

doc-fence-languages.yml, header emptied (71 comment lines to 0, 5559 bytes to 838):

HEAD blob     : 179f1c9eb62c9855aaed84900b22c2f115d67647
mutated blob  : e7e0c9e33d3277950b85b794949c5a1913b60a24   -- differs from the HEAD blob, so the mutation landed
vitest exit   : 1
 FAIL  scripts/__tests__/check-doc-fence-languages.test.ts > its header states the population and never counts it
AssertionError: doc-fence-languages.yml: the header prose this pin reads came back empty or near-empty,
                so it asserted over nothing: expected 0 to be greater than 400
 Tests  1 failed | 27 passed (28)
restored blob : 179f1c9eb62c9855aaed84900b22c2f115d67647   == HEAD blob, and `git diff HEAD` empty

doc-component-types.yml, header cut to two comment lines (68 to 2, 5304 bytes to 824):

HEAD blob     : 18e8bbdbb004b0275a139fcb2dbde0c157a59bcf
mutated blob  : 8710adabc5b7d47be8c874e92a7e4211ca9ed63e   -- differs from the HEAD blob, so the mutation landed
vitest exit   : 1
 FAIL  scripts/__tests__/check-doc-component-types.test.ts > its header states the population and never counts it
AssertionError: doc-component-types.yml: the header prose this pin reads came back empty or near-empty,
                so it asserted over nothing: expected 62 to be greater than 400
 Tests  1 failed | 55 passed (56)
restored blob : 18e8bbdbb004b0275a139fcb2dbde0c157a59bcf   == HEAD blob, and `git diff HEAD` empty

Exactly one test failed in each run, and it named itself and its surface. The near-empty leg is the one that matters: 62 to be greater than 400 is the floor firing, not an emptiness check.

Restore used git checkout HEAD -- ABSOLUTE_PATH and was proven by state — blob back to the HEAD blob and git diff HEAD empty for that path — never by the exit code of the restore itself. A restore step's failure mode is exit 0 having done nothing.

The three copies still hold one pattern

objectui#7943 proved this a few hours ago and the card requires it re-run after this change:

$ grep -h -A1 '^const POPULATION_COUNT' THE_THREE_PIN_FILES | grep -v '^const' | grep -v '^--' | sed 's/^ *//' | sort | uniq -c
      3 (the single regex literal, one distinct line)

$ grep -h 'text.matchAll(new RegExp(POPULATION_COUNT.source' THE_THREE_PIN_FILES | sed 's/^ *//' | sort | uniq -c
      3 (the single documentCounts body, one distinct line)

One distinct line, count 3, for the pattern and for the function that applies it. No divergence introduced.

---

2. The census of content/docs/** — NON-ZERO, so no expansion here

The half that could not be inherited from PR objectui#7915. That census covered 34 .github/workflows/*.yml headers; this card's whole subject is that no copy of the pin reads the published docs tree at all, so that half had to be measured fresh.

Corpus, derived rather than hard-coded — the same derivation the third copy already uses for scope:

apps/site/source.config.ts   dir: '../../content/docs'   ->   content/docs
membership : git ls-files (tracked only), filtered to .md / .mdx
size       : 184 file(s) · 35194 line(s) · 1270087 byte(s)
unit       : one LINE, matching how all three pins read a header (line-filtered, then joined)
pattern    : the family pattern, verbatim from all three copies — a numeral with optional
             thousands separators, then ZERO TO TWO intervening words, then one of
             `.mdx` / `.md` / documents / pages / docs / files, with a negative lookbehind
             ruling out a preceding hash, word character or dot

Hits — the family pattern fires: 6

#	site	matched text	verdict
1	content/docs/guide/ci-cd-pipeline.md:576	18 guide files	historical, measurement point declared. The sentence opens "Measured on main@6422aa891". The gate prints 20 guide file(s) today, so the figure has moved — but the sentence stays true, because it says which tree it measured.
2	content/docs/guide/ci-cd-pipeline.md:781	143 files	live claim, no measurement point. "measured across 143 files and 558 literals". The gate's own verdict line prints 889 type literal(s) today: the companion figure has drifted 558 to 889.
3	content/docs/guide/ci-cd-pipeline.md:1083	207 files	live, drifted, and asserts currency — "the other four still return 203 of today's 207 files". check:shell-escape-residue prints 206 file(s) ... examined in total today. The strongest member: the word "today's" is the claim rotting out loud.
4	content/docs/guide/ci-cd-pipeline.md:1440	73 files	drifted 73 to 107 (scripts/__tests__/*.test.ts), but a test-file population, not a document population. See the proposal below.
5	content/docs/guide/ci-cd-pipeline.md:1659	19 files	historical, window declared — "Across all 5281 first-parent commits on main, 12 commits modified a pre-existing changeset (19 files)". Same class as #1.
6	content/docs/fields/file.mdx:89	3 files	quoted rendered output, inside a fence: // Renders: "3 files" or "document.pdf". The PM ruling of 2026-09-06 applies verbatim — this family governs prose assertions about a document population, not quoted program output.

By segment: prose 5 · fenced code 1 · frontmatter 0.

Blind-spot reading: 18

Required by the card, and the reason a zero would have been believable. The definition mirrors the family's own pattern shape rather than a loose window — each category is a way a real count can be present and still be unreadable to this pattern:

id	blind spot	count
B1	file unreadable / undecodable (any read throw, or U+FFFD after a UTF-8 decode)	0
B2	numeral, then three to six intervening words, then the noun — the family allows at most two	2
B3	the noun before the numeral — the family reads one direction only	12
B4	the count spelled as a word rather than a numeral	4
	blind spot total	18

B1 being zero is the load-bearing one: every file in the corpus was read and decoded, so the 6 above is a reading over the whole tree and not over the part that happened to parse.

B3 is dominated by fenced configuration (max_files: 5, pageSize?: number; // Items per page (default: 10)) which the prose/fence split already separates. B4 contains one genuine word-spelled count — content/docs/guide/ci-cd-pipeline.md:304, "by which time five files had accumulated the defect".

An earlier, deliberately over-inclusive pass (any numeral with a population noun within 60 characters) returned 112. The excess over 18 was ordered-list markers (1. Reads ...), issue references, TS error codes, version numbers and the English quantifiers one / every / all / both. Both readings are recorded here so the tightening is visible rather than assumed.

Ruling: report-only

Census 6, blind spot 18. Neither is zero, so the coverage expansion does not land here. Nothing was added to any exemption table — an exemption table records adjudicated exceptions, it is not a switch for turning a first run green.

---

3. Proposal for the expansion card

The expansion needs three criteria that the workflow-header version never needed, because docs prose is a different register. Each is measured above rather than anticipated:

1. Prose versus fence. A workflow header has no code fences, so no copy of the pin has ever needed this distinction. On a docs page, quoted program output lives in fences — member Fix documentation deployment for www.objectui.org #6 is the instance. The prose/fence/frontmatter split used in this census is the minimum shape.
2. A declared measurement point makes a count non-rotting. Members Implement visual designer for Object UI schemas #1 and Implement component reordering via drag-and-drop in designer canvas #5 name the tree they measured (a commit sha; an explicit commit window). Both are stale against today and both are still true. A pattern that cannot see the difference would red two correct sentences, which is how gates get deleted.
3. The noun set is wider than the family's stated subject. files? catches test-file and source-file populations, not only document populations — member Add default props to all components to prevent collapse in designer #4 is the instance. Either narrow the noun set on the docs surface, or accept that the docs census governs all file populations and say so in the header.

Members #2, #3 and #4 are live drifted counts on content/docs/guide/ci-cd-pipeline.md and must be cleared before the expansion can land, or the expansion reds on its first run. They are filed separately rather than ridden along here: the card's file surface is these two test files, and content/docs/ is read-only for it.

Filed out of this run, all three unassigned for PM triage:

• objectui#7965 — the three live drifted counts on content/docs/guide/ci-cd-pipeline.md, each measured against its own gate's verdict line. Also records the two neighbouring sentences that are stale-looking and correct, so nobody rewrites them.
• objectui#7966 — the coverage expansion itself, Blocked-by: objectui#7965, carrying the three criteria above and the blind-spot note.
• objectui#7967 — a fourth divergence inside this family, found while writing the floor and not repaired here. Both twins keep the # markers the third copy strips, so a count wrapping across a comment line is invisible to both and visible to the third. Measured: markers kept scores [] on a wrapped 184 / pages, markers stripped scores ["184 pages"]. Latent — all three headers are clean under both extractions today.

---

Verification

All from f949c4381, the final commit on this branch, in the worktree /home/user/objectui-issue-7901. Every heavy run went through the shared verify lock; every exit code was captured after redirecting to a file, never through a pipe.

command	exit	evidence
pnpm exec vitest run scripts/__tests__	0	RUN v4.1.10 /home/user/objectui-issue-7901 · Test Files 107 passed (107) · Tests 3224 passed (3224)
the two pin files alone, before this change	0	Test Files 2 passed (2) · Tests 82 passed (82)
the two pin files alone, after	0	Test Files 2 passed (2) · Tests 84 passed (84) — the two new controls
pnpm type-check:scripts	0	tsc -p tsconfig.scripts.json, no output
pnpm lint:root	0	32 problems (0 errors, 32 warnings), all pre-existing; neither edited file appears
pnpm check:control-bytes	0	OK (scanned 6414 tracked text file(s); skipped 85 binary)
pnpm check:vi-mock-specifiers	0	OK (4339 tracked source file(s), 2616 test-named ...)
pnpm check:vi-mock-inherit	0	OK (... 225 call site(s) judged ...)
pnpm type-check:coverage	0	45/46 via type-check · 42/42 packages compile their tests
pnpm lint:coverage	0	46/46 packages linted, 0 with outstanding errors
pnpm check:unreferenced-sources	0	Every shipped source file in every covered package is reachable.

The vitest banner reads RUN v4.1.10 /home/user/objectui-issue-7901 — the repository root, not a package directory (objectui#3378), and no argument was passed after a bare double dash (objectui#3288).

Changeset. node scripts/check-changeset-presence.mjs, verbatim: ✅ No source or published contract of a released package changed in this range, so no changeset is owed. Two test files, nothing published. None added, not even an empty declaration.

Session: https://claude.ai/code/session_01KbJQ1y1J12nZxYzFWhP8Q3

🤖 Generated with Claude Code

https://claude.ai/code/session_01KbJQ1y1J12nZxYzFWhP8Q3

---

Generated by Claude Code

---

Generated by Claude Code