docs(eval): v6 results + the pr_review→enforce promotion-criterion ADR

[williamzujkowski]

Part of #3845. Depends on #3847, #3848, #3840 (ladder ADR's evidence-threshold schema).

Context

Run the panel over the n≥50 set; publish v6 (per-voter tables, deltas vs v5, honest FP analysis under the rubric). Define the promotion criterion per the Epic D evidence schema (e.g. sustained precision ≥ X over a sliding window of N live advisory reviews + eval ≥ Y) and encode it in the claims registry so claims:check tracks it.

Acceptance criteria

• v6 doc with per-voter precision/recall
• Promotion-criterion ADR (schema-conformant; ratification path = Epic D vote)
• Claims registry: '100% bug-catch (v5, n=10)' superseded by the v6 entry

Evidence required

v6 artifact; registry diff; docs gates.

Out of scope

The promotion itself (a future ratified Epic D transition).

[williamzujkowski]

Readiness assessment (2026-06-21 backlog deep-dive). The infrastructure deps are done — #3847 (curation pipeline to n≥50) and #3848 (per-voter precision/recall in the outcome store) are both CLOSED, and the labeling rubric is committed (docs/research/pr-review-eval-labeling-rubric.md). But this issue is BLOCKED on a real eval run, not on writing:

• No committed labeled eval-set data file, no v5 results doc, and no v6 numbers exist in-tree.
• The "v6 results" half requires actually running pr_review over the curated n≥50 labeled set with live adapters and capturing real precision/recall — which must not be fabricated (CLAUDE.md: never fabricate eval/soak data; BLOCKED over a fake pass).

Recommended split to unblock the actionable half:

1. #3849a — run the v6 eval (cost-gated, live adapters): execute the curated eval, commit the labeled set + v6 results doc. The blocker.
2. #3849b — promotion-criterion ADR: encode when pr_review earns enforce (the precision/recall/n floors, grounded in ADR-0017's advisory→enforce floor). Draftable, but its thresholds should be justified by (1)'s real numbers, so it should follow (1).

Until a real v6 run lands, leaving this open and blocked is correct. Not picking it up autonomously to avoid fabricating results. Trigger to unblock: a controlled live-adapter eval run is scheduled/authorized.