process-gap(implementer+adversary): fix-phase adjudication labels (F-CRIT-001, F-P2-MOD-02) leak as comment provenance in production code

## Summary

Implementer agents cite fix-phase adjudication labels (e.g., `F-CRIT-001`, `F-P2-MOD-02`, `F-HIGH-3`, `F-HIGH-4`) as comment provenance in production code. These labels are adversary-pass finding IDs from earlier convergence cycles; they do not resolve to any external artifact after convergence closes.

## Evidence from ftc-blue W3B2

STORY-5.05.005 `scripts/systems/puzzle_system.gd`:
- Line 116: `# Reference: STORY-5.05.005 v1.5 Implementation Guidance / F-CRIT-001.`
- Line 141: `# F-HIGH-3: second ERROR log...`
- Line 128, 317: `F-HIGH-4`
- Line 303: `# F-P2-MOD-02: mirror enter_puzzle's puzzle_states guard...`

Flagged as F-P10-STORY-5.05.005-L03 [pending intent verification] by the Pass 10 adversary. Reader tracing `F-P2-MOD-02` finds no downstream lookup path.

## Adjudication needed

Two options:
- **KEEP (traceability breadcrumb):** Rewrite as proper cross-ref (e.g., `See adversary-pass-02.md F-MOD-02`) so future maintainers CAN look them up.
- **SCRUB (convergence-cleanup):** Delete fix-phase labels at convergence; keep only prose rationale.

Related distinction: L-W3B2-05 already prohibits fix-phase labels leaking as BC anchors (e.g., `See BC-x §X (Path B)`). This issue extends the concern to fix-phase labels as comment provenance more broadly.

## Requested fix

1. Author a convention decision (KEEP or SCRUB) at the engine level.
2. Update implementer prompt template with the decision.
3. Add adversary rubric check for the chosen convention.
4. Optionally: language-specific linter rule (gdlint / clippy / eslint plugin).

## Cross-refs

- L-W3B2-05 (fix-phase labels as BC anchors — already codified)
- F-P10-STORY-5.05.005-L03 (Pass 10 finding, ftc-blue `.factory/cycles/cycle-001/STORY-5.05.005/adversary-pass-10.md`)