fix(reward-memory): hint reviewed feedback ingestion on Lark inbox drain

[huangruiteng]

Summary

Add an advisory feedback-review hint to non-empty, registry-routed Lark inbox drains when the same registered agent has an eligible Reward Memory route.

中文：补齐“读到反馈 → agent 审核提炼 → 显式预览/写入 → 读回”的提示链路，不自动采集群聊，不新增偏好账本，不改变 ACK 或权限。

Root cause and prior work

• automatic_ingest=false disables module-owned automatic hooks, not explicit ingest-event.
• feat(reward-memory): add scoped feedback adapter #2189 supplied the generic scoped-feedback adapter. Issue Fix has explicit ingestion and recall callsites, but no general inbox learning hint.
• feat(reward-memory): recall guidance before outbound messages #3968 wired outbound recall only.
• feat(reward): project inbox feedback into operating lessons #1988 was closed unmerged because its second operating-lesson ledger was too broad. This change reuses the existing candidate/standing-policy/ingest/readback lifecycle.

Behavior and boundaries

• Existing reward_memory capability owns the effect-free hint; the bundled Lark extension is wired at the CLI composition root.
• Exact Goal/Agent/registry command binding; only active writable scoped-feedback routes with enabled standing policy and exact peer identity are listed.
• No provider calls, message parsing, candidate creation, new authority, automatic flag changes, or settlement gates.
• Agent verifies source authority and current evidence, previews a compact event, and checks exact readback after any authorized execution. Allowed actor roles are constraints, not sender verification.
• Disabled/empty/unconfigured/invalid/incompatible routes and explicit config/project overrides preserve prior output.
• JSON and Markdown both expose the guidance.

Validation

• 92 focused tests passed across feedback hints, experiment routing, ingestion, outbound guidance and routed Lark inbox tests.
• Generated preview command exercised through the real CLI: automatic-ingest-off preview succeeds; cross-agent candidate is guard-blocked.
• Read-only local runtime drain verified hint presence while automatic ingestion remained off; no memory provider calls or writes. No live memory ingested.
• Ruff and git diff --check passed.
• LoopX check: five-file public boundary scan clean; zero errors, three pre-existing unrelated registry/history warnings.
• Full canary and remote CI not yet run at PR creation.

Scope and risk

Five cohesive files: product helper/wiring, canonical capability/inbox documentation and regression tests. No private chat, local config, runtime state, credentials or internal evidence included.

Future-facing pass: reuse the existing scoped ingestion owner; no new store, queue, scheduler or hook framework is warranted. The only new behavior is opt-in advisory output, not enforced learning. Runtime change requires separate review; this PR does not install or self-merge.

[huangruiteng]

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机

这个 PR 补上了一个真实但此前断开的操作链：Agent 从 registry-routed Lark inbox 读到用户反馈后，系统没有告诉它如何在既有 Reward Memory 权限和 route 内进行“人工审核、压缩、预览、显式写入、精确读回”。本改动只在 Reward Memory 已对同一 Agent 启用、inbox 非空、并存在 active/writable scoped_feedback route 时给出提示，动机成立。

改动思路

实现把 hint 放在 Reward Memory capability 内，由 Lark CLI composition root 在 drain 后调用；它复用现有 experiment resolver、surface route、scoped_feedback adapter 和 reward-memory ingest-event，没有再造第二套账本、队列或 ingestion authority。这个归属和复用边界是合理的。

我重点核对了默认关闭隔离、authority 语义和 guidance/obligation 区分：hint 明确标为 advisory_only，不会调用 provider、不会自动生成 candidate、不会授予新权限，也不会阻塞 reply/material-review/ACK。automatic_ingest=false 仅关闭自动 hook，不会错误禁止已有的显式 ingest 路径；这个行为变化在 PR 描述与文档中均有披露。

具体改动

• 新增 effect-free build_feedback_review_hint，只投影 exact Goal/Agent/registry 绑定、可写 active route 和 standing-policy 约束。
• 仅对非空、registry-routed 的 lark-inbox drain 注入 reward_memory_feedback_review；empty/disabled/unconfigured/explicit override 路径保持原输出。
• JSON 与 Markdown 都展示 advisory、preview command 和 route 约束。
• 文档说明 source/authority/freshness/conflict 检查、compact event、preview-before-execute、exact readback 和 no-memory 合法结果。
• 新测试覆盖 enable/disable、跨 Agent、wrong peer、unsupported adapter、policy disabled、read-only、真实 CLI preview、scope expansion guard、JSON/Markdown rendering 和 drain parity。

本地复核：96 passed in 13.26s（feedback hint、experiment routing、scoped feedback、outbound guidance、routed Lark inbox），相关文件 ruff check 通过，git diff --check 通过；远端 14 个 checks 均为 success/skipped-as-designed。

对主干的风险

风险低到中等：这是一个 opt-in operator-facing 输出变化，不改变 inbox 内容、ACK 状态或 memory 写入效果。主要风险是提示被误读为授权或自动学习；当前 typed fields、instruction、route eligibility 和负向 parity 测试已经把该风险约束得比较充分。preview command 暴露 invoked registry 路径，但该输出本来就是本地 operator inbox surface，且路径用于精确绑定同一 registry，没有扩大 public/provider 边界。

默认关闭证明覆盖了共享 drain 输出面和显式 override；authority 不是通过 prose 猜测，而是继续由现有 typed route/standing-policy/ingestion guards 执行。未发现会阻塞合并的问题。

我的整体评价

正向，建议合并。它修复了“看到反馈但不知道如何走既有安全写入链路”的可用性缺口，同时保持了默认关闭、显式执行、最小 authority 和精确读回。未来导向检查也没有发现需要在本 PR 中追加的抽象：继续复用 Reward Memory 的单一 ingestion owner 比新增 inbox-specific learner 更合适。中文 README 的对应操作说明可以后续补齐，但英文 canonical capability 文档和 Lark inbox 文档已足以描述本次行为，不构成本次 blocker。

English verdict: No blocking findings. The change adds a narrowly scoped, advisory-only bridge from a non-empty registry-routed Lark inbox drain to the existing explicit Reward Memory ingestion workflow. It preserves feature-off parity, does not call providers or mutate inbox/memory state, does not grant authority or create a settlement obligation, and relies on existing typed route and ingestion guards. Focused local validation passed (96 tests, Ruff, and diff check), and all remote checks are green. Approve for merge.