fix(quota): qualify explicit selection before final guard

[huangruiteng]

Summary

Follow-up to #4332, rebased against main at 0933d48efc058d9e1585bd30058942416aa258a9.

The original seven-file head is mostly superseded by #4335 and conflicts with the current quota transport. The remaining incremental bug is narrower: when a Todo is added after the first guard and explicit selection then hits a final workspace/scope boundary guard, qualification was skipped and the CLI degraded to quota_unexpected_collection_error.

This PR keeps the existing TypeScript qualification and Python quota/receipt owners, computes qualification before returning the final guarded route, and adds one real CLI regression. Deferred selection remains typed (quota_action_selection_deferred, control_repair), the identity-less receipt is replayed without mutation, and retry re-enters the current guard. No new schema, authority, quota spend, provider, or capability activation is introduced.

Validation

• python -m pytest -q tests/control_plane/test_quota_settlement_cli.py::test_selection_added_after_pending_guard_reports_final_boundary — passed.
• Focused neighboring workspace/monitor/replan/user-gate cases — 5 passed.
• python -m pytest -q tests/control_plane/test_quota_settlement_cli.py — 58 passed.
• python -m pytest -q tests/control_plane/test_quota_boundary_projection_selection.py tests/control_plane/test_cli_output_budget.py — 25 passed.
• Ruff, Python compile, and diff checks — passed.
• loopx canary premerge --from-git-diff --git-diff-base origin/main — direct checks and most selected smokes passed. quota-plan-smoke reproduces an existing capability-memory effect-runtime fixture failure on both baseline and this branch; catalog-run-e2e-smoke passes when invoked directly with the project Python but is reported failed by the canary wrapper. These are environment/baseline gaps, not failures in the changed path.

UI impact: none. This is a quota CLI/control-plane behavior fix.

[huangruiteng]

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

动机

评审 exact head：4350@3f1a7c0bd94aa9d7dfc1bfdb049250d372944c82。本 PR 从 #4332 的旧七文件批次中收敛出当前 main 仍缺失的真实边界：首轮 quota guard 后新增一个显式 Todo，随后该 Todo 因最终 workspace/scope guard 暂不可交付。修复前 should_run_packet 在 guarded-agent-lane 分支提前返回，CLI 丢失 action_selection_qualification，最终退化为 quota_unexpected_collection_error。这会掩盖可恢复的 control_repair，让 Agent 不能按同一 Turn 可靠重入。

改动思路

入口是 quota should-run --codex-app --todo-id。TypeScript action-selection reducer 仍是资格语义 owner，should_run_packet.py 负责从 canonical Todo inventory 组装 route，quota.py 负责 CLI 错误与 heartbeat receipt。修复只把现有 qualify_action_selection_from_inventory 调用置于 final guarded-route return 之前；selected identity 仍由 guard 冻结，资格仍是 derived projection。deferred/rejected 继续走现有 typed preflight，已有 identity-less receipt 只 replay，不创建 settlement identity、不写新 rollout event、不扣 quota。没有新增模块、schema、authority、provider 或 capability activation。

具体改动

关键代码讲解

• loopx/control_plane/quota/should_run_packet.py::_resolve_agent_lane_delivery_route：先计算请求 Todo 的 qualification，再返回 workspace/scope guarded route；非 guard 的现有 qualified/candidate 路径复用保存的 qualification，避免重复计算。
• loopx/cli_commands/quota.py::_apply_requested_quota_action_selection_preflight：未改动，继续作为 deferred/rejected 到 CLI/receipt 的唯一 transport owner；本 PR 验证它现在能收到 final-guard 的 typed qualification。
• loopx/control_plane/work_items/action_portfolio.py::qualify_action_selection_from_inventory：未改动，继续从 canonical inventory 推导 task class 并调用 TS reducer。
• tests/control_plane/test_quota_settlement_cli.py::test_selection_added_after_pending_guard_reports_final_boundary：新增 real File-authority CLI 回归，覆盖首轮 identity-less receipt、Todo late add、control_repair typed defer、receipt replay/no mutation、确定性 retry 和最终 guard flags。

对主干的风险

这是局部 ordering 修复，运行时 diff 仅 8 行，覆盖现有 explicit-selection 调用，不触及默认选择、monitor polling、autonomous replan、user gate、workspace repair 已有路径或 quota settlement。负向 counterexample 在 baseline 上复现了 quota_unexpected_collection_error；exact head 返回 quota_action_selection_deferred，reason=control_repair，receipt status=replayed，rollout event 数量保持不变。新测试使用真实 quota CLI、File canonical authority 与持久化 receipt readback，不依赖 mock postcondition。

验证结果：focused regression 1 passed；workspace/monitor/replan/user-gate 邻近场景共 5 passed；完整 test_quota_settlement_cli.py 58 passed；boundary/CLI 两个相关 suite 25 passed；Ruff、Python compile、diff checks 通过。loopx canary premerge --from-git-diff --git-diff-base origin/main 的 direct checks 和大多数 selected smokes 通过；quota-plan-smoke 在 baseline 与 head 都受 capability-memory effect-runtime fixture 影响，catalog-run-e2e-smoke 使用项目 Python 直接运行通过但 wrapper 报失败，属于已复现的环境/基线证据，GitHub CI 仍是最终门禁。代码和测试均不改变 public/private boundary。

我的整体评价

当前 two-commit exact head 已完成 #4332 的 rebase 增量收敛：保留唯一仍有价值的 final-guard qualification 修复，删除并行 transport、重复协议文档和重复 monitor 测试。repository reuse、typed state、authority semantics、default-off isolation、domain neutrality 与 guidance-vs-obligation 均通过；没有 blocking finding。由于这是 author-owned PR，GitHub 不允许正式 self-approve，以下 COMMENTED 结论等价记录本次 APPROVE；合并前仍需 exact-head readiness 为 ready 且 GitHub 必需 checks 通过。

English verdict: APPROVE for 4350@3f1a7c0bd94aa9d7dfc1bfdb049250d372944c82 (author-owned PR; recorded as a comment because GitHub blocks self-approval). The focused ordering fix preserves typed quota_action_selection_deferred/control_repair and receipt replay for late explicit selections under the final boundary guard. Existing TS/Python owners, defaults, authority, and settlement paths remain unchanged. Focused, full quota, boundary/CLI, Ruff, compile, and diff checks pass; two canary failures are reproducible baseline/environment gaps and require GitHub CI confirmation.