duplicate-decl-gate misses two #1485 variants: deep Markdown headings, and duplicate YAML keys inside fenced blocks in ci-standards.md (fleet-propagating)

[don-petry]

Follow-up to #1025 / #1033. Not a defect in #1033 — that gate is a clear improvement and catches the recurring shell-function class cleanly. This records two variants it does not catch, so the gap is not assumed closed.

How this was found

The prior session's handoff recorded that #1033's gate, executed against #1024's corrupted content, blocked "both the shell duplicate and the AGENTS.md heading duplication." Re-running the gate at #1033's head 47778d92 against #1024's actual tree (663a334a), only the shell duplicate is reported:

::error::Duplicate top-level declarations found — the #1485 corruption class.
- `./scripts/compliance-audit.sh` — duplicated top-level function declarations:
  - `ensure_required_labels` (declared 2 times)

Running it against only AGENTS.md + standards/ci-standards.md from that same tree reports nothing, while both files demonstrably carry duplicates.

Gap 1 — duplicated deep (####) Markdown headings are not detected

#1024's AGENTS.md carries #### Pull Request Limits (automation open-PR cap) at both line 599 and line 1091. The gate reports nothing; it appears to consider only top-level headings.

Note the constraint that makes this non-trivial: repeated headings under different parents are legitimate Markdown and are live on main today — AGENTS.md ### Agentic Directives ×3 and standards/ci-standards.md #### Adopting in a new repo ×2. The gate correctly passes main today, and must keep doing so. So the fix is not "flag all repeated headings" — it needs to key on the heading's full ancestor path, or on adjacency/near-duplication of the section body.

Gap 2 — duplicate YAML keys inside fenced Markdown blocks (highest blast radius)

#1024's standards/ci-standards.md emits the claude-issue: job key twice (lines 926 and 961), both inside the ```yaml fence opened at line 865. Fenced blocks are deliberately skipped — correct for prose examples — so this can never be caught by the current design.

This is the variant that matters most: standards/ci-standards.md is the template adopting repos copy their workflows from. A duplicate YAML mapping key makes the resulting workflow unparseable, so a corruption here propagates fleet-wide as broken CI in every repo that syncs the stub. It is the highest-blast-radius form of the #1485 class and it is currently unguarded.

Acceptance criteria

• Duplicated Markdown headings are detected at every depth, keyed on the full ancestor path so sibling sections under different parents are not flagged.
• .github main and .github-private main still pass — verify explicitly against AGENTS.md ### Agentic Directives ×3 and ci-standards.md #### Adopting in a new repo ×2. A gate that turns main red is worse than no gate.
• Duplicate mapping keys inside fenced ```yaml / ```yml blocks in Markdown are detected, at least for repeated keys at the same indentation within one fence. Prose fences (bash, text, untagged) stay exempt.
• Regression tests using feat: implement issue #1023 — canary autocut: scope bump signals to watched-path commits, paginate the range, and persist failed promotions #1024's real content as fixtures: AGENTS.md (#### Pull Request Limits ×2) and ci-standards.md (claude-issue: ×2) must both fail the gate; the current main copies of both files must pass.
• Consider folding in the repeated-block detection from Two duplicated blocks (#1485 class) are live on .github main — deploy-standard-workflows.sh and compliance-retrigger.sh; the sweep only catches duplicated functions #1040 (≥3 significant repeated lines) rather than building a third separate gate.

Not in scope

Widening to arbitrary structural duplication. The three concrete shapes actually observed in this corruption class are: duplicated shell functions (caught today), duplicated Markdown sections, and duplicated YAML keys.

Related: #1025, #1033, #1024, #1039, #1040, #1485, petry-projects/.github-private#1520, petry-projects/.github-private#1609.

[don-petry]

This scope gap has now cost us three live duplications — evidence

The standing sweep run on 2026-09-13 against petry-projects/.github-private main (ca733ebc) found three #1485 duplications that the shipped gate cannot see. Filed for removal as petry-projects/.github-private#1779.

The root cause is one line — scripts/check-duplicate-decls.sh:

done < <(find "$SCAN_DIR" -name '*.sh' -type f | LC_ALL=C sort)

-name '*.sh', SCAN_DIR defaulting to scripts/. Markdown and YAML are not merely under-checked for the two variants this issue names — they are never opened at all. The gate reported green on every PR that carried these files.

What it missed:

file	defect	introduced by	live since
prompts/dev-lead/fix-ci.md	Phases 2/3/4 duplicated, 46 lines byte-identical	a3d903e4 (#1519)	2026-08
prompts/dev-lead/fix-reviews.md	Phases 2/3 duplicated, 20 lines byte-identical	a3d903e4 (#1519)	2026-08
personas/dev-lead/persona.yml	reusable: twice in the runtime: mapping	d10483a5 (#1697)	2026-09-08

Two details worth carrying into the fix:

1. a3d903e4 is titled "fix: restore trunk from #1485 re-corruption". The restore commit restored the corrupted copy of both prompts. A gate that only sees *.sh cannot catch a bad restore of a *.md, and a human reading that commit title reasonably assumes trunk is clean. That is precisely the failure mode this issue exists to close.
2. d10483a5 rode into ring0/ring1. The persona.yml duplicate was merged on 09-08 and promoted onto the canary tiers before anyone looked. The gate ran green on PR #1697.

These are the prompts dev-lead actually executes, so items 1 and 2 are behavioural, not cosmetic — every fix-ci and fix-reviews run has been reading its phase list twice for weeks, while petry-projects/.github-private#1741 is open about dev-lead exhausting its action budget.

One constraint on the widening

Whatever this issue ships must not fail main. The same sweep produced these hits, and all of them are correct code:

• AGENTS.md — ### Agentic Directives ×3, and standards/ci-standards.md — #### Adopting in a new repo ×2. Different parent sections. Already pre-registered in Two duplicated blocks (#1485 class) are live on .github main — deploy-standard-workflows.sh and compliance-retrigger.sh; the sweep only catches duplicated functions #1040.
• tests/dev-lead/fixtures/conflict-integrity/*.sh in .github-private — intentionally corrupt fixtures that the gate's own bats tests assert against. Scanning tests/ naively will fail the gate on its own fixtures.
• Repeated owner: / issue_number: / labels: keys across with: blocks of different steps in the same workflow file — same indent, different mappings. A duplicate-YAML-key check that keys on indent alone reports ~20 false positives per repo; it has to track mapping scope.
• Keys and headings inside block scalars (e.g. Glob: lines in the prompt body of feature-ideation-reusable.yml) — must not be parsed as YAML or Markdown structure.

Recommend the acceptance criteria here require a regression test that runs the widened gate over current main of both repos and passes, alongside the tests that fail on the #1779 pre-fix content.