M5-3f precision docs frozen: 61/61 — the judge and the independent audit agree on every answered pair

The five t3-precision docs, generated at the previous commit and
gated from this one on (T-G1..T-G14 all live in CI):

- Overall precision 1.000 (61 correct, 0 wrong) against the 0.85
  contract; per-corpus 1.000 everywhere the answered denominator
  reaches 5 (cobra 12, requests 11, ripgrep 18, zod 19; self answers
  1 — published, no gate, the small-denominator stance).
- The separation is the finding: every audited boilerplate row (16
  in zod alone — the vitest test()/schema/expect scaffold family the
  calibration warned about) scored BELOW the 0.85 TSED threshold and
  landed in not_clone; the judged-clone set drew exclusively from
  the copy-lineage truths {clone, variant, t1t2}. The θ table
  publishes the full grid; 85 is the loosest grid threshold clearing
  the contract on this sample (the 80 and 75 columns start admitting
  boilerplate).
- Universe ledgers ride each doc from a full product run on the
  pinned tree: reported pairs 219/842/84/3773/10036 against the
  pre-registered floors 53/63/18/617/3253 (T-G14, all clear) plus
  the over-cap and forest drop ledgers. The forest cause is now in
  the frozen record twice over: the audit's unit_gaps rows show
  anonymous-closure units sliced as whole lines and chain fragments
  — spans that cover sibling nodes have no single root (the
  short/anonymous-unit alignment thread, already on the ledger).
- min_answered=40 was met by the mains alone (61 judged clone), so
  no backup top-up ran; the mechanism stays documented in the method.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: skymanbp