[codex] #88 支持 dogfood answer facets

[catoncat]

Mainline Intent

Intent: int_3266cd23
Status: proposed
Title: 支持 dogfood answer facets

What changed

为私有 dogfood acceptance flow 增加 expected.answerFacets、facet/assertion 分离评分、失败分类、runner/README/scorecard 输出，以及 passing/failing facet 测试覆盖。

Why

#88 要求 dogfood case 不只验证 session/cwd/snippet，还要能验证检索到的 evidence 是否支撑真实回答 facets，并把 failure 归类成可行动的问题层级。

Decisions

• answer facet schema 位置: 在 expected.answerFacets 下加入 label、mustContain、可选 failureClass (answer facet 是对 retrieved evidence 的预期，不是 find option；放在 expected 可与现有 session/cwd/context assertions 共存。)
• facet 和既有断言如何评分: DogfoodEvaluation 增加 assertionMark、facetMark、failureClasses，同时保留总 mark/blocking (runner 可分别报告 facet pass/fail 和 session/cwd/match/context assertions，同时 hard/candidate gate 仍使用总体 mark。)
• failure classification 的表达: 新增枚举 coverage_index、skill_guidance、cli_recall_ranking_context、stale_golden、unclear_case，并附在失败 predicate/facet 上 (Answer-facet dogfood gate v1 #88 需要失败能落到可行动层级；默认 recall/context 问题归 cli_recall_ranking_context，无命中归 coverage_index，facet 可显式覆盖。)
• session-level evidence read: runner read-range 对 matchSeq 缺失的 hit 使用 --query 回退，并使用 sessionRef 读跨 source session (answer facets 必须绑定真实 evidence；session-level hits 不能因为缺 seq 而跳过上下文读取。)
• review 后修复: 抽出 buildDogfoodScoreboard，并让 README predicate 输出显示 answer_facet: (独立 Standards review 指出 scoreboard 重复和 answer_facet.answer_facet 可读性差；已修。)

Subsystems: eval, dogfood, acceptance-gate, agent-recall, perf

---

Summary by cubic

Adds answer facet verification to dogfood evals and acceptance, with separate scoring for assertions vs evidence-backed facets, plus failure classification and updated scoreboards. Satisfies #88 by checking that retrieved evidence supports real answer facets and by grouping failures into actionable classes.

• New Features
  • Schema: add expected.answerFacets[] with label, mustContain[], optional failureClass.
  • Scoring: compute assertionMark and facetMark in addition to overall mark; keep hard/candidate blocking on overall mark.
  • Failure classes: support coverage_index, skill_guidance, cli_recall_ranking_context, stale_golden, unclear_case; aggregate to failureClasses.
  • Scoreboard: new assertionPass/Fail and facetPass/Fail; buildDogfoodScoreboard used in acceptance gate and runner; perf bench parser updated.
  • Output: README/scorecard show assertion vs facet marks, failure classes, and predicate labels like answer_facet:<label>.
  • Context read: desiredContextMode triggers on facets; for read-range fallback to --query when matchSeq is missing; use sessionRef for read-range and read-page.
  • Tests: added passing/failing facet cases and acceptance-gate coverage.

^{Written for commit a3c6517. Summary will update on new commits.}