fix(audit): honesty fixes from benchmark run 1

[DocRoms]

Findings du benchmark A/B 0.8.13

Le run 1 du gate pré-MEP a attrapé deux défauts d'honnêteté réels : la mécanique terminale (Interrupted, zéro baseline, zéro validation) a par ailleurs tenu parfaitement.

• Placeholder detector : le code markdown est invisible au compteur : l'index TD documentait {{DO_NOT_1}}/{{DO_NOT_2}} en code inline et un fichier de 43 TDs a échoué le step 8 comme « still the template ». Strippers anti_halluc partagés, durcis fail-closed (5 passes de review adversariale) : spans inline appariés par runs de N backticks (run non fermé = texte gardé), fences lexicales strictes (opener ≥3 backticks / ≤3 espaces / info-string sans backtick ; closer même char / run ≥ / suffixe espaces-tabs only ; non fermé = tout restauré, ligne d'ouverture incluse).
• Compteurs de réconciliation persistés sur Interrupted : resolved/new étaient calculés puis jetés : « 0 new » s'affichait comme valeur connue alors qu'un TD réellement nouveau existait. La partition exacte atterrit désormais en DB.

Tests

• Backend 3562 (dont 8 pins nouveaux : repros exacts du banc + faux opener / faux closer / markdown tronqué) · clippy 0 · double review Claude+Codex en room

Gate : après merge → rebuild → session MCP neuve (bridge_info stale=false) → rerun benchmark v3.7 complet.

[Copilot]

Pull request overview

This PR fixes two “honesty” issues surfaced by benchmark run 1: (1) placeholder detection falsely counted placeholders that were merely documented inside Markdown code (inline/fenced), and (2) Interrupted audit runs failed to persist reconciliation counters (resolved/new), causing the UI to show misleading “0 new” values.

Changes:

• Persist td_resolved_since_last and td_new_since_last alongside the other TD counters for Interrupted runs via update_td_counts, and update the Interrupted finalize path to pass these values.
• Harden Markdown fenced-code stripping to be lexical and fail-closed (unclosed fences restore withheld content) and expose the stripping helpers for reuse within the crate.
• Update placeholder counting to run on text with fenced blocks and inline code spans removed, and add targeted regression tests for the benchmark failure modes.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated no comments.

File	Description
backend/src/db/audit_runs.rs	Extends update_td_counts to persist reconciliation counters and adds/updates tests for Interrupted runs.
backend/src/core/anti_halluc.rs	Reworks fenced-code stripping to strict lexical rules with fail-closed semantics; exposes stripping helpers; adds regression tests.
backend/src/api/audit/validation.rs	Applies the shared Markdown strippers before counting raw {{IDENT}} placeholders; adds regression tests to ensure code spans/fences don’t count.
backend/src/api/audit/full.rs	Updates the Interrupted terminal write path to pass resolved/new/carried into update_td_counts.