fix(extraction): #885 the pre-judge ceiling now names the candidates it cut

[jasonssdev]

Closes #885

The defect

The 24-candidate pre-judge ceiling was the only drop in the extraction pipeline that reported a bare count with no titles. In the wild run that surfaced this, every other drop named names — 4 wrong-language, 9 recombined, 10 judge, 1 staging — and this one said 7.

That contradicts the principle the module already wrote for its sibling in _unevidenced_notice:

"The titles are load-bearing, not decoration. A notice that only counted would leave the operator opening every derived document to find the one."

You cannot check against the source what you cannot name.

After

merged extraction union exceeded the 24-candidate pre-judge ceiling;
1 merged candidate(s) never reached the judge: Subject 25

Three deliberate choices

• A separate field, not a replacement. pre_judge_dropped_titles sits beside the int. Stored runs predating it carry () and still render the half they always had; replacing the count would break reading historical data for a cosmetic gain.
• Additive wording. The count and the ceiling explanation stay. That line is what an operator reads to understand why extraction stopped short — trading information for information is not an improvement.
• The siblings' shape. Same _CAP_NOTICE_TITLE_LIMIT + (+N more) as every other title-bearing notice, so the extraction notices read alike.

What this does NOT fix

#885 has two halves and this closes one. Truncation is still positional — merged[:_MAX_JUDGE_CANDIDATES] over a list concatenated in window (document) order — so the tail of a long source is what goes, regardless of quality. Naming the cut makes that observable, which is the step before fixing it. Both the field docstring and the test record it, so the next person to touch this does not have to rediscover it.

The substantive remedy — ordering candidates by something other than document position before the cut — needs measurement first and is left open.

Evidence

• RED first: three new notice tests failed on the missing field; the fourth (a stored run with a count and no titles) passed from the start, pinning backward compatibility.
• The end-to-end test was extended, not duplicated. It already builds 25 distinct candidates so the ceiling cuts exactly 1, and now asserts the notice names Subject 25. Mutating the implementation to take merged[:n] (the head) instead of merged[n:] (the tail) turns it red — so the assertion pins which candidate is dropped, not merely that some title appears.
• Full suite 5658 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