Feedback UI for per-sentence model advice: synthetic demo block, counts first, registration identity

[Jeffreyliu0131]

Feedback UI line of the model-reception wave (task book 143, line 149; merges in the same batch as 148 or right after it). The v4 feedback page section "理由与证据的支持关系" now shows 148's per-sentence relation advice, which closes 148's user-visible gap: until this line, a synthetic registration showed "模型建议失败". The 2026-10-12 decisions apply here:

• A synthetic registration shows labels and evidence, and the whole block is marked "合成机制演示".
• The UI leads with sentences. After review 156 the user confirmed that the folded first screen shows per-sentence counts first and the work-level conclusion after, in lighter text.
• The read side re-checks the work-level conclusion against the conservative aggregation rule.

Per-criterion advice for the 14 items (UI-03) is out of scope by the 10-12 cut.

What changes

• Reading (readRegisteredAdvice, signature unchanged, optional fields only):
  • It reads units, aggregation, and the identity fields kind, dataset, components, fusion and weights_stage.
  • Labels and evidence are accepted only when status=synthetic_mechanism_only and the registration scope is synthetic_fixture. An external registration with a synthetic status, or a synthetic registration with completed, is rejected.
  • Invariants:
    • Failed or unavailable advice may carry no label, at work or sentence level.
    • Counts must match the units: labels counted by label, abstentions and failures by status.
    • The work-level label must equal a conservative re-aggregation of the sentences.
    • Units and counts come as a pair, and sentence indexes are unique.
    • Every reference belongs to this session.
    • Fusion members are in fixed left/right order with weights left 1−alpha and right alpha. Without a fusion record, member weights must sum to 1.
    • alpha and the threshold lie in [0, 1].
    • A criterion task, criteria, rubric_version or reconciliation is rejected.
  • A violation drops only that item to unavailable/model_prediction_invalid; other advice, rule results and the stored feedback are untouched. Read objects carry no request_id, job_id, input_hash or unit_hash.
• Folded first screen (updateRegisteredAdvice, signature unchanged):
  • Lines: title, synthetic notice, non-zero per-sentence counts, then the work-level conclusion in lighter text. Example: "合成机制演示 / 合成或未训练权重，不是模型判断，不计分。/ 逐句：支持 1 · 矛盾 1 · 证据不足 4 / 作品整体：存在矛盾". In English: "Synthetic mechanism demo / … / By sentence: supported 1 · contradicted 1 · insufficient evidence 4 / Whole work: contradiction found".
  • Title priority: "模型建议失败" (load failure) → "模型调用超时" → "模型未判断任何句子" → "合成机制演示" → "模型建议（不计分）" (external candidate) → the existing waiting and pending states.
• Expanded view:
  • Failure reason and "no advice adopted" first.
  • An ordered sentence list: quote truncated at 120 characters, result, and evidence buttons labelled with object and version ("材料 policy · 第 1 版", replacing the 64-hex evidence id; the id stays in data-model-reference). An evidence button opens the source in the side panel.
  • The note "按规则切句，不是陈述抽取", plus skipped-short and over-budget counts.
  • Registration identity: model, kind, version, training data with label tiers (e.g. "G2v 模型复核标签", never "human labels"), weights stage, fusion members with weights, alpha, and whether dev showed a fusion gain. External candidates add "外部候选；模型质量尚未验证".
• Failure states (load failure, timeout, nothing judged) show no label and no evidence button, and say "未采用任何模型建议；原作品与已核实的规则结果都保留".
• "评价模型：…" line above the block: read only from the JUDGE slot recorded in this feedback's provenance ("未接入" when unconfigured, "未记录" when missing); the support model is added only when configured separately; no model name is hard-coded.
• Copy: all module copy lives in the paired table MODEL_ADVICE_COPY (model-advice-copy.ts), with identical zh/en keys and placeholders. Focus is restored per item, sentence, evidence and reference. Only existing v4 color variables are used, dark mode follows automatically, and there is no motion.
• MODEL-02 browser check now expects "Synthetic mechanism demo" and the notice instead of "Mechanism validation"; the three protected meanings are kept: not shown as a model judgment, folded by default, zero calls.
• Files only under apps/web/src/features/feedback-native/. No backend, contract, 144 or 148 file changed. contracts-v2.ts is byte-identical to the generator output for manifest bdeee6db….

Review and verification

• Independent review 080-review-149.md (156) on 592cab8: 可合入, no P1 or P2, 9 P3.
  • The reviewer walked the normal entry in zh and en with a synthetic linear registration, and with a 145 fusion at alpha 0.75 on a tree merged with 148 ebc6238; the member weights 0.25 / 0.75 displayed correctly.
  • 25 hand-built invalid inputs each fell back on their own.
  • 18 mutations all went red.
  • The merge preview with 148 ebc6238 was conflict-free.
• After the review the user confirmed P3-1 (counts first, conclusion after and lighter), P3-2 (re-aggregate the work-level label on read) and P3-3 (member weights sum to 1 without a fusion record). These landed in 2e56be1, the final head; no separate re-check of that commit is recorded.
• 50 new test IDs: v4-slot.test.ts 24, model-advice.test.ts 12, model-advice.browser.mjs 14. MODEL-02 was updated. Reds were taken on the base (node-base-red-full.txt, browser-base-red.txt, p3-base-red.txt), and 22 mutations on the final code all went red.
• Line checks on 2e56be1:
  • Targeted backend: 525 passed.
  • Frontend vitest: 440 passed / 7 skipped / 0 failed; against the base's 411 front-end IDs, 0 missing, 0 regressed, 36 new.
  • 72 state checks; build; lint 0 new.
  • Module browser check: 27 passed.
  • tsc --noEmit passes.
• Normal serve + worker + v4 run (2e56be1, synthetic linear registration, zh and en, 18 screenshots):
  • The folded screen reads as above. The rule range line "规则已定标部分：0–45 分（适用项共 45 分，满分 100）" is unchanged.
  • Six sentences expand with the expected results, and the policy evidence button opens "差旅住宿与交通政策 · v1" (en "Travel policy: hotels and taxis · v1").
  • No horizontal scroll at 390px.
  • HTTP re-reads were byte-identical, and the 12 journal files were unchanged.
• The controlled page (26 screenshots) covers six states: external, fusion, synthetic, load failure, timeout, all abstained. Each is shown in zh and en, desktop light and 390 dark, plus keyboard shots.
• Full gate ran once on the integration tree 11e53dd7… (146 b765c88 → 145 23fb9df → 147 1e2b85c → 148 261118e → 149 2e56be1, integration branch HEAD 855b366; 157 receipt).

Not claimed / registered

• UI-03 (14-item advice and conflict display) not built, by the 10-12 scope cut. A criterion entry would currently read as invalid and turn the block title into "模型建议失败" (P3-4). This is unreachable today because criterion_advice is always None; when §8 resumes, skip such entries instead.
• Shown only on the controlled page, built by changing states on recorded 148 output with identities from 145/146 registrations:
  • external candidate;
  • fusion and encoder identity lines;
  • load failure, timeout and all-abstained states.
• The normal-entry registration is a W01 linear bundle that publishes only four legacy identity fields, so kind, training data and components do not appear there. Normal-entry walks of fusion, encoder, drift, and the evaluation model under a loopback JUDGE slot (FLOW-03/04/05) belong to 150.
• Real model quality not verified; only synthetic registrations were used.
• P3 registered:
  • docs/design/frontend-style.md §2.1 forbids model IDs and "training" wording, which conflicts with the required identity display; 150 registers the exception.
  • Copy details: the English "every sentence supported" overstates when short sentences were skipped; n=1 plurals; only 3 abstention codes are mapped.
  • Evidence buttons show object ids, and config items open with no body text.
  • The coverage line still shows a percentage, contrary to 10-12 item 3; this belongs to W05, not this line.
• Not owned here: docs/modules/expansion-v3/W08.md still says "机制验证" and hard-codes port 19810. The replacement text is in the receipt, for 150.

Receipt: 完成回执/claude-coordination/149-反馈界面模型建议显示.md.