fix(#3812): document how Current Position actually resolves a duplicate field (#4017)

* docs(#3812): say that Current Position is single-valued, and pin the behavior that makes it true

#3812 shipped CLOSED with half its acceptance unmet. #3873 delivered cardinality for FRONTMATTER
keys - current_phase/current_plan render as optional at docs/reference/state-md.md:89,91, covered by
tests/gen-state-md-docs.test.cjs:374. The issue's actual ask was the ## Current Position BODY
section, and that never landed. Surfaced by an /adr-phase-coverage audit of epic #3473; the issue
was reopened rather than noted.

The section now states three things: every field is single-valued, the section is overwritten rather
than appended to, and a duplicate resolves to the FIRST occurrence with no warning - so a line
appended in good faith is silently ignored rather than winning. Progress history belongs in
## Performance Metrics, two headings down, and the text now points there.

The third claim is a behavioral promise about the reader, so it was VERIFIED BY EXECUTION before
being written rather than inferred from the issue title:

  stateExtractField(<"Phase: 1 of 5 (First)" ... "Phase: 9 of 9 (Appended later)">, "Phase")
    -> "1 of 5 (First)"

The mechanism is state-document.cjs:405 - the plain-line pattern ^<field>:[ \t]*(.+) carries flags
im with NO g, so String.match returns the first hit. Writing "first wins" without running it would
have repeated the exact error I had to retract twice in this epic already.

A test pins the reader, not the prose. Three rows in tests/state.test.cjs: T1 (load-bearing) asserts
the duplicated case resolves first; T2 asserts the ordinary single-field case still works, so a fix
that only functions when duplicated cannot pass; T3 puts a Plan: line BETWEEN the two Phase: lines
and asserts it resolves independently - negative space, because a reader returning the first line of
the SECTION rather than the first matching FIELD would satisfy T1 alone. Proven to discriminate: a
last-match variant returns "9 of 9 (Appended later)" and T1 reds.

No assertion checks that the document contains a sentence. That is what local/no-source-grep exists
to stop, and it would pin wording that is allowed to improve. The point of the test is that if that
regex ever gains g and a last-match walk, the test fails - instead of the documentation quietly
becoming a lie with nothing to notice.

Prose only, no new heading. docs-state-md-locale-parity compares heading-level sequences by LCS
rather than text, so added paragraphs cannot fail it while an added HEADING would fail all four
locales. The constraint is structural, not stylistic - confirmed by running that comparison after
the edit.

The four locale copies are translated rather than left stale. They are not gate-enforced for prose,
so "nothing fails" was available and is not the same as correct: leaving four documents asserting
something the English one now contradicts is a correctness problem. Code spans and the anchor link
stay untranslated - they name real tokens.

The whole approach rests on one fact, checked first: ## Current Position at :196-208 sits OUTSIDE
every generated marker region (:81-104, :138-151), so a hand edit survives --write. Re-confirmed
after all five edits - gen-state-md-docs --check reports all 6 targets up to date. Had that been
false the fix would have belonged in the generator, and a hand edit would have been silently
reverted.

One real gate failure fixed inline rather than reported: the new test's comments referenced
docs/reference/state-md.md, which was not in that file's registered exempt-docs paths, and
lint-docs-guard-registration failed lint:ci correctly. Registered.

Known limit, named rather than folded in: gsd-tools validate/health still do NOT warn on a
duplicated Phase:. #3812 records that as a "consider", not a requirement, and confirms none of the
nine rules in src/health-diagnostic-rules/{state-consistency,phase-structure}.cts counts
occurrences. Documenting the silent first-match is the delivered scope; making it loud is new scope
and stays unclaimed.

Closes #3812

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* docs(#3812): the rule I documented was false — replace it with the measured one

An isolated review returned two blockers. Both mine, and the first is the worse
kind: I wrote a falsifiable rule into a reference page and got it wrong.

1. "A duplicate resolves to the FIRST occurrence" is FALSE.

   stateExtractField (src/state-document.cts:401-419) tries BOLD `**F:**` across
   the whole input, THEN plain `^F:`, THEN a pipe-table row. Form precedence
   beats document order. Measured against the built reader, all intra-section:

     Phase: A (plain)  /  **Phase:** B (bold, later)   -> B    LATER WINS
       Phase: A (indented) / Phase: B (plain, later)   -> B    LATER WINS
     Phase: A (plain)  /  | Phase | T (table) |        -> A    first wins

   My original verification tested plain-versus-plain, saw first-wins, and
   generalized to all forms. Measuring one case and claiming the general rule is
   the same error I have had to retract twice already in this epic.

   It is also worse than silence. The sentence told authors an appended line is
   safely ignored; a bold line appended "for emphasis" silently overrides the
   original. Someone trusting the doc would have corrupted their own state file.
   And #3812 never asked for a resolution rule - it asked for single-valued,
   overwrite-not-append, and where history goes. The rule was my unrequested
   addition.

   Replaced with the measured truth: resolution is by FORM (bold anywhere, then
   plain at line-start, then table row), and only WITHIN the winning form does
   the first occurrence win. Both consequences stated plainly - a higher-ranked
   form wins regardless of position, and an indented `Phase:` is invisible to the
   plain form. All five claims in the new paragraph verified by execution before
   being written, including the two I had wrong.

2. The tests tested the wrong case and passed for the wrong reason.

   T1/T3 put the second `Phase:` under `## Somewhere else` - the INTER-section
   case, which #2956 already fixed by scoping. #3812 says verbatim that #2956
   "fixed the inter-section case and never addressed intra-section duplication",
   so the case the new prose describes was untested, and the fixtures passed
   because of section scoping rather than field resolution. They also called bare
   stateExtractField rather than the production chain, T2 could not discriminate
   first from last at all, and no fixture mixed forms - which is precisely why the
   false claim survived to review.

   Rewritten as four rows, all intra-section, all through the real
   stateCurrentPositionSlice -> stateExtractField path: plain-then-plain (first
   wins within a form), plain-then-bold (the bold LATER value wins - the row whose
   absence let the false claim ship), indented-then-plain (indented invisible),
   and sibling-field independence. Each proven to fail against a reader that
   disagrees.

3. Two dead anchors. pt-BR and zh-CN linked `#performance-metrics` while their own
   headings are `### Métricas de Desempenho` and `### 性能指标`. Both fixed to the
   anchor their own heading generates. ja-JP/ko-KR kept the English heading, so
   theirs already resolved.

4. A ja/ko sentence inverted its own meaning. Both rendered "which is the section
   designed to grow" with a bare demonstrative whose nearest referent read as
   Current Position - saying the opposite of the point. Rewritten so the clause
   attaches unambiguously to `## Performance Metrics`.

5. Cross-locale drift, flagged by the implementing agent rather than by me: after
   fixing EN, the four locales still stated the OLD false rule. Four documents
   asserting something measured to be wrong is worse than four saying nothing.
   All four now carry a faithful translation of the corrected paragraph, with
   code spans, each file's own anchor, and the ja/ko referent fix preserved.

Verified: all five claims executed against the built reader; every rewritten test
row proven to discriminate; gen-state-md-docs --check reports all 6 targets up to
date, so the edits stay outside the generated marker regions; locale heading
parity unaffected (prose only, no headings added); build:lib, lint and lint:ci all
exit 0.

Refs #3812

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* docs(#3812): second false rule on the same page — scope the ranking to the section

A second isolated review found a second false falsifiable claim, and the failure
mode is the same one twice in a row:

  attempt 1: verified plain-vs-plain, wrote a claim about ALL FORMS
  attempt 2: verified bare stateExtractField, wrote a claim about THE DOCUMENT

Both times the claim covered a wider surface than what was actually executed. The
fix each time was not a better sentence, it was executing the surface the
sentence describes.

BLOCKER — "bold `**Phase:**` anywhere in the DOCUMENT wins" is false.

  ## Current Position / Phase: 1 of 5   +   ## Archive / **Phase:** 88

    bare stateExtractField(whole doc) -> "88 (other section)"
    PRODUCTION (slice then extract)   -> "1 of 5 (in section)"

#2956's section slice means production never hands another section to the
matcher; a bold line in `## Archive`, or in the YAML frontmatter, is simply not
seen. The ranking is real but scoped: it applies WITHIN `## Current Position`.
I verified against the bare function and wrote a claim about the system.

Every existing test placed its bold line inside the section, which is exactly why
nothing contradicted the claim. T5 now puts a bold `**Phase:**` in `## Archive`
and asserts production returns the in-section plain value, with the unscoped
reader asserted to DISAGREE so the row proves the scoping rather than assuming
it.

BLOCKER — the changeset still shipped the ORIGINAL retracted claim.

I corrected the page and left the release note saying "resolves to the first
occurrence ... a second entry added in good faith is silently ignored". The note
contradicted the page it announces, and the release note is what most people
actually read. Rewritten to the corrected rule.

MEDIUM — the concession was inverted. It read "wins even if it comes FIRST in the
file", which is the vacuous direction; the surprising case, and the one the very
next clause illustrates with an APPENDED bold line, is "even if it comes LAST".
All four locales reproduced the inversion faithfully, so it was an EN-source
defect rather than translation drift.

Two sharp edges now named, both measured: a bold `**Phase:**` followed only by
trailing spaces resolves to an EMPTY STRING and does not fall through to a valid
plain line below (T6 pins it); and `| **Phase:** | 3 of 4 |` short-circuits to the
bold form and returns the literal `"| 3 of 4 |"`. A page that teaches form
ranking has to say where the ranking bites.

Also fixed: all five files labelled the link `## Performance Metrics` while the
heading is `### Performance Metrics`. Anchors resolved correctly everywhere; only
the label's level was wrong.

Every clause in the final paragraph re-verified through the PRODUCTION chain
(stateCurrentPositionSlice -> stateExtractField), clause by clause, before being
written: bold in another section does not win; bold in frontmatter does not win;
bold appended last does win; first wins within one form; trailing-space bold
yields empty. All four locales carry the same corrected rule.

gen-state-md-docs --check reports all 6 targets up to date; heading counts
unchanged at 20/20 across all five files, so locale heading-parity is untouched;
build:lib, lint and lint:ci all exit 0.

Refs #3812

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* chore(#3812): backfill changeset pr number

Refs #3812

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

---------

Co-authored-by: sim <sim@local>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Tom Boucher
2026-08-28 20:50:15 -04:00
committed by GitHub
parent 1051c6d8d4
commit ac7587287b
8 changed files with 182 additions and 2 deletions

View File

@@ -196,6 +196,10 @@ Phase-completion verbs never write `Milestone complete` (the overloaded bare val
| `Last activity:` | 핸들러가 기록할 때 ISO 날짜(`YYYY-MM-DD`); 실행기가 작성할 때 서술형 산문 |
| `Progress:` | 시각적 막대, 예: `[████░░░░░░] 40%` |
**이 섹션의 모든 필드는 단일 값이며, 이 섹션은 추가되는 것이 아니라 덮어써집니다.** 두 번째 `Phase:` 줄은 두 번째 위치가 아니며 이력도 아닙니다 — 이는 잘못된 형식의 입력입니다. 리더는 중복을 문서 내 등장 순서로 해석하지 않습니다. **형식(form)**에 따라 해석합니다. 판정 순서는 각 형식이 `## Current Position` **섹션 내** 어디에 나타나든 관계없이, 섹션 내 어디에든 있는 굵은 글씨 `**Phase:** value`, 그다음 줄 시작에 오는 일반 텍스트 `Phase: value`, 마지막으로 파이프 테이블의 `| Phase | value |` 행 순입니다 — 그리고 승리한 형식 **내에서만** **첫 번째** 항목이 우선합니다. 이 섹션 밖에 있는 굵은 글씨나 일반 텍스트 줄 — 이전의 `## Archive` 항목이나 YAML 프런트매터의 굵은 글씨 줄 — 은 전혀 참조되지 않습니다. 실제 구현은 항상 먼저 `## Current Position` 섹션으로 범위를 좁힌 뒤(#2956) 형식 우선순위를 적용합니다. 실무적으로 두 가지 결과가 따릅니다. 상위 형식으로 작성된 중복은 섹션 내에서 **나중에** 등장하더라도 승리합니다. 즉 "강조를 위해" 뒤에 추가된 굵은 글씨 줄은 무시되는 대신 앞선 일반 텍스트 줄을 조용히 덮어씁니다. 또한 들여쓰기된 `Phase:` 줄은 (줄 시작을 기준으로 삼는) 일반 텍스트 형식에는 보이지 않으며, 다음으로 일치하는 형식으로 넘어갑니다. "상위 형식이 이긴다"는 규칙에는 두 가지 주의할 점이 있습니다. 값이 후행 공백뿐인 굵은 글씨 줄은 아래의 유효한 일반 텍스트 줄로 넘어가지 않고 빈 문자열로 해석됩니다. 또한 레이블 셀 자체가 굵게 표시된 파이프 테이블 행(`| **Phase:** | value |`)은 굵은 글씨 패턴에 먼저 포착되어 파이프를 포함한 셀 원문이 그대로 반환됩니다. 이 섹션을 작성할 때는 줄을 추가하지 말고 교체하세요.
진행 *이력*은 여기에 속하지 않습니다. 이는 계속 늘어나도록 설계된 섹션인 아래의 [`### Performance Metrics`](#performance-metrics)에 누적됩니다.
이 섹션의 `Status:` 및 `Last activity:` 필드는 기존 값이 알려진 템플릿 기본값인 경우 GSD 핸들러에 의해 업데이트됩니다(크누스 불변량: 실행기가 작성한 값은 보존됩니다). 알려진 핸들러 기본값의 전체 목록은 `gsd-core/bin/lib/state-document.cjs`의 `KNOWN_TEMPLATE_DEFAULTS`에 있습니다.
### Performance Metrics