Bind reader-appropriateness to scoring, revision and delivery (#592)

[Stahl-G]

First, contained slice of the #592 plan (private_planning/issue-592-delivery-consistency-plan-2026-09-10.md). It makes the writing-noise problem observable to users without adding roles or a new revision loop.

What changed

• Gate decoupled from input field. release.decision decides soft/hard from the saved reader-contract clause type (requirement_severity), so the same sentence typed into the objective or into writing preferences cannot flip a formal-delivery gate.
• Score-to-revision connected. Expression <=2 is must_fix; it triggers the existing single auto-revision even when the overall verdict would look passing. A must-fix body marked 达到要求 is recorded as assessment_inconsistent and shown in the assessment panel ("表达分偏低但总体仍判为达到要求，两者不一致…"), not only left in the events table. Expression anchors 4/3/2/1 are in the evaluator rules.
• One effective rule set. report_profiles keeps only the industry increment; shared body/record, accuracy and evaluation rules stay in deliverable_spec.
• Handoff. generation_prompt writes scout-contract.md and hands every Scout its absolute path plus the saved plan.json reader_contract through the real dispatch; the scout contract text points at plan.json instead of implying it is self-contained. build_packet refuses to drop a saved reader contract.
• Gaps carry impact. Delivery-affecting gaps record related/impact/action/status; execution detail stays out of the body.

Tests

• Behaviour checks: gate is independent of the input field; must_fix/overall_inconsistent thresholds; generation_prompt writes the scout contract and wires it into the dispatch.
• pytest: 183 passed, 3 subtests with node_modules present. Without it, test_rich_document cannot import @tiptap/core (environmental, unrelated to this diff).

Not in this PR

• Clause-level tracking (4.2) and the rich-text must-fix highlight (4.6).
• Structuring gaps into a Pydantic shape (review fix: use github.event.before for README check base ref #3) — changes the data shape, separate PR.
• The real closed-loop acceptance run.

Ref #592.

[cursor]

Review

上一轮要的三件事都做了：表达分与总体矛盾进评分面板、Scout 交接有行为测试、requirement_severity 用 SOFT_CONTRACT_KINDS。门不再看输入框、表达 ≤2 会走现有一次修订、行业增量与共用规则拆开，方向对。CI 绿。可以合；下面几处是切片里还虚的地方，不否定 #592 这一刀。

已对齐

• 正式交付 soft/hard 跟已保存条款类型走；无合同时才回退 kind == 'writing'。
• must_fix（表达 ≤2）触发 auto_revise；达到要求 + 表达 ≤2 记 assessment_inconsistent，面板有说明，不替模型改分。
• generation_prompt 写出 scout-contract.md，payload 带绝对路径，提示里点名 plan.json 的 reader_contract。
• 缺口四字段写在共用 instructions 里，不改数据形状（和「Not in this PR」一致）。

建议收紧

1. build_packet 的约定丢失检查在生产路径上基本走不到。 _snapshot 已经是 resolve(requirements, reader_contract=detail.get('reader_contract'))，再比较 detail.reader_contract 和 requirements.reader_contract 等于用同一来源和自己比。以后真的漏拷贝，也要先改 _snapshot 才会碰这行。要么在测试里塞一份被改过的 snapshot 证明它会响，要么断言「detail 有约定则 spec 必须带上」，不要只做恒真比较。

2. 门控测试没有覆盖「同一个句子换输入栏」。 test_writing_constraint_gate_ignores_which_input_field_it_came_from 只有一条 objective，把整段（含「说明客户交付变化」）标成 writing_preference 或 reader_content。soft 用例等于把必答内容和写作约束绑在同一个 id 上再整段变软——这是 Orchestrator 漏拆时的失败态，不是「写进 writing_preferences 栏」的成功态。真正的栏位无关应是：内容句在 objective、写作句在 writing_preferences（或都在 objective 但拆成两条 clause），未覆盖写作只出 notice、未覆盖内容仍 blocker。

3. Scout 交接仍靠提示词。 文件写了、prompt 里有路径，测试也只断言这两点。scout_slots 没有结构化带上 scout_contract_path / plan.json。模型漏贴路径时，Python 不会知道。本切片可以先这样；下一刀最好把路径写进 slot，而不是只写进 Orchestrator 说明。

小点

• 评分面板用 Number(d.expression)<=2。expression 为 null 时 Number(null)===0，会误出「表达分偏低但总体达到要求」。Python 的 must_fix 要求 isinstance(score, int)，两边不一致。用 Number.isInteger(d.expression) && d.expression<=2。
• 行业提示里删掉的「不为披露缺口加分」「正式交付资格由独立检查决定」在 deliverable_spec 里只有部分对应；若还靠行业增量承担，确认一下不要只靠共用读者规则兜。

结论： 作为 #592 第一刀可以合。优先补一门真正换栏位的测试，以及别让 Number(null) 在面板上误报。

[cursor]

只有一条 objective，把整段标成 writing_preference / reader_content。soft 用例等于内容句和写作句绑在同一个 id 上再整段变软，测的是漏拆，不是「同一句换输入栏」。应覆盖：内容在 objective、写作在 writing_preferences（或同栏但两条 clause），未覆盖写作只 notice、未覆盖内容仍 blocker。

[cursor]

_snapshot 已经用 detail.reader_contract 去做 resolve，这里再和自己比，生产路径上几乎不会响。要证明「丢掉就失败」，测试里需要一份被改过的 snapshot，或断言 detail 有约定则 spec 必须带上。

[cursor]

Number(null)===0，缺 expression 且 overall 为「达到要求」时会误出这段。Python must_fix 要求整数分。用 Number.isInteger(d.expression) && d.expression<=2。

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.