C03

The requested correction belongs in flag-presence matching, with coupled tests and documentation. Static inspection establishes the failure; no tests or workflows were run.

Current behavior

src/gzkit/doc_coverage/flag_scanner.py:220–259 checks if flag in content. Therefore a manpage containing only --output-format satisfies coverage for --output. Presence reads the entire UTF-8 manpage, permits mentions in examples/prose, maps spaces in command names to hyphens, skips absent files, and returns existing {path, issue} dictionaries. Explicit waivers override the default empty _PER_FLAG_DOC_WAIVERS.

The same module already has a reusable boundary rule: _mentions uses re.escape(flag) followed by _FLAG_END = r"(?![\w-])" (lines 281–291). Truth checking uses that rule to avoid --attestor/--attestor-present collisions. It checks required-ness and value-taking only within fenced Usage/Synopsis blocks. The actual adr-emit-receipt.md independently names both siblings.

commands/cli_audit.py:242–274 derives presence names and truth specs from one scan, appends both issue lists, sets top-level valid, prints human/JSON output, and exits 1 on findings. quality.py:798 runs this audit; the command quality registry includes it. Command-level cross_coverage and the separate doc-coverage chore report do not include per-flag issues. A flag finding can therefore coexist with passing command-level coverage without requiring a schema change.

Necessary coupled obligations and boundaries

Reuse the existing suffix-aware matcher for presence, adapting its usage-only documentation to describe arbitrary text. Preserve whole-page presence scanning; borrowing the truth check’s Usage-only scope would reject currently valid prose/example documentation.
Update the coverage docstring’s explicit substring contract and the Presence description in docs/user/manpages/cli-audit.md to explain that longer siblings do not count. Correct any real manpages newly exposed by the stricter check; their identity/count needs a later census. Do not add waivers simply to restore green results.
Preserve waiver lookup, missing-file delegation, slug/path construction, issue wording and ordering, absent-source fallback, and the shared AST discovery. This request does not require a parser/discovery rewrite, new dependency, output-field change, or expansion of the chore report.
Preserve truth checks when sharing the matcher. Its current suffix rule rejects continuations containing letters, digits, underscores, and hyphens, while permitting whitespace, punctuation, backticks, =, brackets, and end-of-text. It is case-sensitive and regex-escapes names.
The existing matcher has no left boundary: prefix--output or ---output can still match. It also allows a dot after the flag. Rejecting every embedded or unusual option spelling would broaden this request; define that separately before strengthening the shared rule and changing truth behavior.

Meaningful verification proposed

First add a direct coverage regression with both flags registered and only --output-format documented. Require exactly the shorter flag’s missing-documentation issue and the correct path. Then add an independent --output mention and require no issues. This fails under the current implementation.
Table-test longer suffixes, accepted Markdown/prose/example forms, --output=VALUE, end-of-file, punctuation, case sensitivity, and literal regex-special characters. Derive expectations from the intended matching contract rather than copying implementation output.
Exercise explicit/default waivers, multiword command slugs, missing files, and independent commands. Retain truth regressions for sibling flags, optional/required flags, valueless arguments, and claims outside Usage.
Exercise CLI audit with actual collision-containing temporary documentation and nonempty discovered specs: verify exit 1, JSON valid: false, the diagnostic, and human failure output, then successful repair. Existing synchronized-doc fixtures can omit CLI sources, causing the scan to return empty; those alone cannot witness this change.
Run the real-manpage waiver census and relevant audit/quality integration tests after implementation. Existing quality tests verify failure propagation through mocks; they complement, but cannot replace, the collision test.

Uncertainty: no repository-wide impact census, runtime verification, ownership assessment, or governance action was performed. No claim of clean current docs or passing checks is made.

Read inventory: /tmp/gzkit-current-inputs/C03.md; under /tmp/gzkit-current-snapshot/: src/gzkit/doc_coverage/{flag_scanner,scanner,manifest,__init__,runner}.py, src/gzkit/commands/cli_audit.py, excerpts of src/gzkit/{quality.py,commands/quality.py}, tests/commands/{test_cli_audit,test_audit}.py, excerpts of tests/test_quality.py, and docs/user/manpages/{cli-audit,adr-emit-receipt}.md. Targeted rg searches also inspected matching lines under src, tests, features, docs/user, and .gzkit/skills; search hits were not treated as whole-file reads. One combined read output was truncated; the omitted relevant test section was subsequently read.

Call count: 9 shell/read calls, including task read and this sole permitted output write. Assessment complete within the bound; implementation and execution remain intentionally unperformed.