Audit H6: an imperative lives in a comment, not in an expression

[ojassug]

Stacked on #5 (→ #4 → #3). Base is audit-gateway, so this diff shows only the H6 commit. Merge the stack bottom-up; each retargets automatically.

Closes audit H6 — the single largest cause of real code not being optimized. CONSTRAINT_DIRECTIVE_LOST accounted for 24 of 40 fallbacks on the audit's corpus.

The audit's proposed fix would have destroyed real protection

The nine-keyword scan (must, must not, never, always, only if, do not, required, except when, make sure to, critical) is written for natural-language system prompts, and was applied to raw content of every kind. In source, required and critical are ordinary identifiers.

The audit's remedy was "scope directive extraction by content type (prose/markdown/prompt kinds only, not code)". I measured where the dropped directives actually come from, over a frozen 293-file corpus at targetReductionRatio: 0.3:

bucket	from comments / docstrings	from code
Python	16	38 — nearly all logger.critical(...)
TypeScript	38	13 — readonly required?, error-message literals

So neither extreme works. Trusting the check everywhere keeps 51 false positives; the audit's version discards 54 genuine constraints — and specifically the Python docstring case that docs/phase-1d-semantic-gate-disposition.md measured this check to be the only thing catching.

What separates the two populations is not the content type but the region. An instruction to a reader lives in a comment or a docstring; it never lives in an expression.

Part 1 — extraction is scoped to prose regions

extractProseRegions returns whole content for prose content types, and for code returns line comments (//, #, --, *), block comment bodies, and Python docstrings including their interior lines.

Deliberately line-oriented and syntax-approximate rather than lexed: this is a filter on what may raise a constraint, so over-inclusion costs a false positive (the pre-existing behaviour) and under-inclusion costs a missed constraint. Requiring the comment leader at the start of a trimmed line is exactly what excludes logger.critical(exc) while keeping # never call this twice.

Part 2 — retention is checked per item

The check collected every item's directives into one list and tested each against after.items.map(i => i.content).join('\n'). Two problems:

• A directive extracted from item A was satisfied if the string happened to appear anywhere in item B — the check could pass for content that was in fact destroyed.
• A loss anywhere failed the whole run, with no way to say where.

Matching by item id fixes both, and the message now names the item. An item absent from after is skipped, on the same reasoning DriftTracker.findUnwitnessedItems records: selection is not elision, and failing here would make any prunable item carrying an imperative unprunable.

Measured

bucket / route	before	after	delta
python (file)	14.98%	23.14%	+8.16pp
python (stdin)	14.88%	22.66%	+7.78pp
typescript (file)	23.38%	27.33%	+3.95pp

20 of 586 rows changed, and none regressed — no file went from reducing to falling back. Every other bucket is byte-identical.

What deliberately remains

After the change TypeScript has zero remaining code-sourced directives. Every surviving CONSTRAINT_DIRECTIVE_LOST is a genuine imperative in a comment or docstring that an elision would drop, and those files still fall back — correctly, since the run would otherwise silently delete an instruction. That is why the category does not go to zero.

Ordering note for reviewers

This had to land after C1a (#3). The audit observed that this check was "currently the only thing preventing markdown documents from being deleted" — a document survived if its author happened to use one of nine words. Narrowing it first would have widened that data loss.

With C1a in place the drift measurement gate covers markdown on its own merits, and the corpus confirms the handoff: the prose bucket is unchanged at 28/28 fallbacks, now attributed to drift rather than to a coincidence of vocabulary.

Verification

• npm run typecheck, npm run lint — clean
• npx vitest run — 518 passing, 59 files (up from 507)
• New tests verified to fail 9 of 11 against the unfixed code; the 2 that pass are the ones pinning behaviour H6 must preserve

Cumulative across the stack

Python 14.98% → 23.14%, TypeScript 14.00% → 27.33% — while every gate got stricter: max passing symbol loss fell 66.7% → 40% (#4), and the surviving-witness rule (#3) closed whole-document deletion.

Still open

H5 / §3.1 (multi-item ingestion — the highest structural leverage left, and what would make the knapsack reachable), C4, H4, H2, and the M-tier stragglers.

🤖 Generated with Claude Code