fix(ingestion): #884 extraction_notice records every condition, not the strongest

[jasonssdev]

Closes #884

The defect

extraction_notice held one value, so when a run tripped several disclosure conditions the strongest token overwrote the weaker at write time. lint — the audit surface — then rendered Unevidenced objects: and Staging-dropped candidates: as independent sections while the second silently masked the first.

The issue's cheapest proposed remedy was "make lint disclose the masking". Implementing it showed that is not possible: the displaced token is destroyed at ingest, so by the time lint reads the document there is nothing left to disclose. That remedy collapses into this one.

The masking was pervasive, not rare

Making the field plural turned ten existing ingest tests red with this shape:

['judge-selection-unavailable',   'objects-without-evidence'] == 'judge-selection-unavailable'
['sole-object-restates-source',   'objects-without-evidence'] == 'sole-object-restates-source'
['objects-without-evidence', 'candidates-dropped-in-staging'] == 'candidates-dropped-in-staging'

Those fixtures had been tripping two conditions and losing one all along — invisible, because the tests only asserted the winner. This was not a quirk of the E2E corpus.

Also fixed: the in-batch collision is no longer a staging loss

When two candidates of one run slugify alike, the first was already staged, so the content is on disk and the bundle represents the source — the identical reasoning the code already applied to the create-only skip. Counting it also left debt no command could clear, since the redo that marker prescribes (--re-extract) reproduces the same collision deterministically.

Precedence moved rather than disappeared

It was a rule that destroyed data in the frontmatter. It is now a presentation rule in the batch summary, which counts a file once however many conditions it carries. lint enumerates them all.

On disk, a single condition stays a bare scalar — byte-identical to earlier releases. Only a genuine co-occurrence widens to a list, and readers accept both shapes, so existing Sources keep their disclosure.

Two false claims already in the code, corrected

• lint.py stated "disjoint by construction: a Source carries at most one token" — true of the tokens, false of the conditions. That sentence was the justification for the broken design.
• A comment added during this change promised a single condition would write the bytes it always did, while the code wrote a list. The code was changed to match the comment.

Tests

Four tests named ..._outranks_... were renamed and strengthened: after the change they passed while asserting only membership, so their names claimed more than their bodies checked.

The headline lint test uses two Sources with the tokens in opposite order. One alone does not discriminate — whichever token sits at index 0 is still found by a reader inspecting only the first element. Mutating either lint reader to doc.extraction_notices[:1] now fails; with a single fixture, neither did.

Specs

ingestion/spec.md's precedence requirement rewritten, its two precedence scenarios replaced, and two scenarios added: an in-batch collision without a marker, and scalar-on-disk for a single condition.

Full suite 5659 passed, 1 skipped. ruff clean. mypy clean over 272 files.

Found during the 0.2.10 manual E2E.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MZz1zEMqg4X5mDEwwfC6dj