fix(extraction): apply the four-lens findings on the name grounding (#712)

[jasonssdev]

Follow-up to #719. Applies the review findings that PR raised, rather than letting them sit behind an approved receipt.

Two of the three were found independently by the resilience and reliability lenses, which is usually the sign a finding is about the code rather than about taste.

Word boundaries

Grounding was a raw substring test, so a short name was "found" inside any unrelated word containing it — Ana in mañana, Vega in Vegas. That silently un-fires the advisory exactly where it matters most: a short fabricated name is both the easiest to hallucinate and the easiest to hit by accident.

The declared bias table names which false negatives this check accepts. This was not among them — it was a bug.

The label-only exemption was all-or-nothing

all() over every matched label let a single Presenter: line disable the exemption for an entire AMI transcript and re-flag every role-titled participant. It is now half of the distinct labels: a transcript with one longer label still does not state its participants' real names.

The legacy path reported nothing

cli/main.py picks extract_concept_union if union_judge else extract_concept, and only the union path computed the advisory — so with union_judge off, participants were stored and nothing was surfaced. That is the "computed but never read" defect #690 already spent a PR diagnosing, one function over.

Also

• The coverage report's rendered line still read anchor-less discards — the one part of that file a human actually reads.
• _participant_unreadmitted_notice now records why its single-cause claim holds and what would invalidate it: the field is a complement, not a cause, and slice 3's budget lane is exactly the change that would make the wording a lie before any test noticed.

Verification

Each fix mutation-confirmed against its exact target line with __pycache__ purged — reverting any one fails its own test and no other. Suite 4802 → 4806 passed, 1 skipped. ruff, ruff format, mypy clean.

Lineage review-5436f9f54c1f3932, HIGH risk, canonical 4R, approved, receipt bound to this exact candidate, pre-pr gate allow.

Known residue, deliberately not chased here

This round's lenses raised further findings on the fix itself, and they are real:

• the trailing half of the word boundary is not independently exercised (in manana the leading lookbehind already rejects, so (?!\w) is untested)
• the docstring's stated reason for preferring lookarounds over \b is wrong — \b matches beside a colon too
• the set vs list change ("distinct rather than every line") is not observable in its test
• nothing pins the direction the widened exemption must still fire in
• "LABEL-ONLY" is now the wrong vocabulary for a half-threshold

These go to one final PR rather than another review-of-the-fix cycle. The exemption threshold in particular deserves a better rule than "half the distinct labels" — most likely the recurring-speaker notion _transcript_shaped_text already uses, which ignores one-off labels by construction.