docs: per-item drift has nothing to attribute, and §48/§50 are why

[ojassug]

Measured before implementing, and the precondition fails. Recording it rather than building it — the disposition BM25 and MMR got.

Every corpus fallback, classified by its actual fallbackReason

Frozen corpus, 288 files, file route, target 0.3:

cause	count	already attributable?
CONSTRAINT_DIRECTIVE_LOST	29	yes — per item, §47
SEMANTIC_DRIFT_UNMEASURABLE	87	yes — unwitnessedItemIds, §33
input not valid UTF-8	1	correct, unrelated (§35)
SEMANTIC_DRIFT_EXCEEDED	0	— the only code per-item drift would help

On multi-item bundles — src/core (33 items), src/stages, src/gateway, src/adapters, pip's 45 Python files, this repo's 62 TypeScript sources — S_k measures 0.0024–0.0056 against a 0.40 threshold, and does not move at ratios 0.3, 0.5, 0.7 or 0.9. Roughly two orders of magnitude under the gate.

The item was not wrong when written — it was closed by other work

§47 measured drift at 0.4122 on an engine that elided everything it could. §48 gave the stage a token ceiling to stop at, and §50 made the unit it removes a statement rather than a whole function body. Together they cut symbol loss far below the gate.

An open item is a claim about the current build, and it expires like any other. This one expired two releases ago and stayed on the list because nobody re-measured it.

The mechanism is real, and reachable only by asking for it

At --max-drift 0.001 — 400× stricter than default — it fires and behaves exactly as §47 predicted: drift names no item, hasUnattributableError is set, repair is declined, and 7 items that would otherwise have been reverted were not. The bundle fell back whole, and the reported reason was the constraint failure the repair would have fixed.

So the design note is correct about what would happen. What is absent is any default-configured input on which it happens. Built as specced it is ~1,000 lines with no observable effect — the H5 condition.

§51 records what would make it live again, and says to re-measure first: it is one corpus run.

Where the slot went instead

docs/audit-remediation-status.md §7 now points at making the constraint gate finer. That is where the fallbacks actually are — 29 of 29 code-bucket fallbacks — and §50 separately measured the same gate as the reason the better sub-region coverage setting costs two working files. It is the per-item axis with something on it.

628 tests green; documentation-only change.

🤖 Generated with Claude Code