Archive current verification and the failed extra A01 assessment

Why: Keep the latest merged candidate reviewable without turning failed model findings into a release claim. Evidence: Exact d06792c composition, two current browser cases and independent model review are linked and source-labeled. Impact: Documentation only; model quality gate and merge remain closed. Refs: Approved U2/U3 task card
2026-10-13 06:23:53 +00:00
parent 1b296a1773
commit dd3b9da619
3 changed files with 108 additions and 2 deletions
@@ -1,4 +1,4 @@
-> 原件归属：独立审阅者 /root/fast_slow_final_review；原始报告与重放日志目录 `/tmp/fast-slow-u2-u3-20261013/review/`。以下为原报告原文存档；其中相对原始日志路径以该原件目录为基准。未复制评测材料或凭据。
+> 原件归属：独立审阅者 /root/fast_slow_final_review；原始报告与重放日志目录 `/tmp/fast-slow-u2-u3-20261013/review/`。当前代码审阅PASS，模型验收FAIL；未复制评测材料或凭据。

 # Independent fix re-review

@@ -93,3 +93,23 @@ The accompanying fake mechanism stream now chooses a real acceptance-condition q
 **Final code-review verdict: PASS / no remaining identified blocking code findings for the reviewed U2/U3 change sequence**, ending Backend6a2d85c and Frontend3087b06, with unchanged Core6e38574 and UI3b34dfd. This is a code-review verdict, not overall delivery acceptance. The inherited login feature was reviewed only at the merge/startup boundary as recorded above. Root7fd5288/fresh-main4ad5b3c and Documents53a7d37 were supplied as delivery composition metadata; full root build/test/CI and planning closeout are executor gates, not independently asserted by this focused replay.

 **Overall U3 acceptance remains NOT PASSED / pending** on the six paid samples. OriginalA01 failure, revisedA02 semantic/fidelity limitations and revisedA04 budget-stopped final output remain documented without alteration. Additional failed-sample rerun approval has not been received or assumed by this reviewer. No code-review PASS should be used to state model acceptance, exact-head CI or merge approval is already complete.
+
+
+## Post-disconnection recovery review — Backend e69865e11ba3f74887f92c2433b5ec37c881d0db / Core beafd7537a6770bd6888095b48926980fce487d8
+
+Read-only review resumed at the current exact objects. Inspected zero-findings fix62a2ab4, upstream merge7855d82 and fixture adaptatione69865e. Checkout was clean; before/after object IDs stayed identical during all replays.
+
+The zero-accepted-findings failure copy is now truthful: only active findings count as available phase findings; absent active findings yield outcome none, explicit no-usable-findings conclusion, honest remaining work and a new-request suggestion without implying stored results. The exact budget exception remains in Run.Error. This closes the extraA01 zero-findings display finding prospectively; original failed sample remains unchanged.
+
+Merged coexistence is intact: Core retains TaskSnapshot/DocumentOrigin while adding SettingScope. Production app injects plugin schema, Cards, execution budgets, shared authorization, mirroring policy, command wiring and PG event wiring before recovery/public intake. SSE still subscribes before readiness, filters task source visibility on replay/live send, deduplicates seq, and now independently rechecks session validity on the upstream timer. No U2 task/source guard was lost in the merge.
+
+The isolated eval fixture now uses one settings service with trusted SetSystemOverride(config), shared by Manager/Server/plugin host. This is compatible with system-managed models/runtime.mock; it avoids silently stripped user overrides. It does not modify a real account setting or introduce a provider/credential change. Material intake thresholds and maxAttachmentChars remain normal effective scoped settings; original material revision logic consumes the effective cap.
+
+Independent REPLAY (all exit0, all exact before/after SHAs recorded):
+
+- `post-disconnect-sqlite.log`: Backend SQLite, `go test ./internal/runmgr ./internal/store ./internal/cardsvc ./internal/server ./internal/app ./internal/settings -run 'TestFailedTaskDoesNotPromiseAutomaticContinuation|TestTask|TestMaterial|TestCardsAtCursor|TestReviewQueuedNotice|TestSSE|TestMergeUserDottedKeyMaskAndNormalizeStored' -count=1 -v`. Covers zero findings, canonical queue/recovery, checkpoint/source authority, atomic card/event failure, visibility, production startup, session SSE invalidation and managed-setting scope behavior.
+- `post-disconnect-pg.log`: same Backend, isolatedPG, `go test ./internal/runmgr ./internal/store ./internal/cardsvc -run 'TestFailedTaskDoesNotPromiseAutomaticContinuation|TestTask|TestMaterialRecord' -count=1 -v`. Covers transactional projection/source/recovery under PG.
+- `post-disconnect-fixture.log`: same Backend, `go test ./internal/server -tags live -run '^(TestExtraA01LimitsActualRunIDs|TestEvaluationPreservesProductionOutputCap)$' -count=1 -v`. Fake streams/HTTP only, no paid models.
+- `post-disconnect-core.log`: Corebeafd753, `go test ./... -count=1`.
+
+**Current code review: PASS for the inspected U2/U3 fixes and integration boundary; no remaining identified blocking code finding.** This does not certify all unrelated upstream settings/auth behavior, final root CI, deployment or overall model quality. **Overall U3 quality remains FAIL/not passed**, including the explicitly approved extraA01. Source can be prepared as a draft/reviewable candidate; no merge or additional paid sample is justified by these mechanism tests. Raw model evidence and all original failure assessments remain applicable as recorded.
@@ -0,0 +1,83 @@
+> 原件归属：独立审阅者 /root/fast_slow_final_review；原始记录在 `/tmp/fast-slow-u2-u3-20261013/review/`。以下为质量判定，未复制完整评测材料或凭据。
+
+# Independent verification — explicitly approved extra A01
+
+Reviewer `/root/fast_slow_final_review`, 2026-10-13. Read-only inspection and fake-provider tests only. No model calls, credentials, source edits or further sample retries. The original six-sample failures remain recorded.
+
+## Object, authorization and artifacts
+
+The parent records explicit user approval for exactly one additional A01, at most two paid execution runs, reported-token target60000 and no automatic retry. This reviewer did not infer further authorization from leftover budget.
+
+Sample manifest fixes Backend `31efafa4790843b39ea15c4c0cd99e1495d89199`, Dev `316b5a5a56c4cf1201003865448c613cbb2a12e6`; binary SHA256 `1ce5c7961b3544bfe337d4495f3bdded776d8b4afd71afcbe56a35bbfb6cd2d5`. 31efafa's delta from6a2d85c changes only live evaluation fixtures/consumer tests, not production behavior. Original evidence remains in `../model-output/candidate-extra-a01.json` and `../approved-extra-a01-data/provider-records.json`.
+
+## Paid run count and usage — independently reconciled
+
+| Actual RunID | Kind/state | Provider calls | Reported tokens |
+|---|---|---:|---:|
+| run_b05e2ef14df2 | conversation/completed | 2 | 10706+11013=21719 |
+| run_8cc08e1fb3fe | subagent/failed | 2 | 9362+11122=20484 |
+| run_05d9df58ee30 | conversation/completed, test-only notice | 0 | 0 |
+
+Exactly **two paid run IDs**, four real provider requests, **42203 reported tokens**. Provider total equals summed persisted Run.Usage, with usageUnknown=false. The two foreground requests match exact non-process persisted text; the two child requests match tool-call IDs. Mapping is saved in `extra-a01-call-run-map.json`. Provider fingerprint groups=3 is not a paid-run count.
+
+Every actual request's outputCap is4000; neither child call began with a remaining run allowance below4000. After its second request, reported20484 exceeded the20000 run threshold, so further child requests were denied. This sample does not exhibit the old fixture cap override. No final child answer was produced. Observed extra usage is below its separately approved60000 target by17797; that difference is not new-run authorization.
+
+Original batch211974 + separately authorized extra42203 = **254177 reported tokens across both authorizations**. Keep those two budgets distinct; the extra authorization does not rewrite the original240000 batch target.
+
+## Checkpoint rejection and source validation
+
+First tool call `call_0VotHi8NLwSGVP42wwmeLq2K`:
+
+- Valid JSON syntax, invalid candidate schema: findings contains unknown `$text`, source fields are misplaced at finding level, one source has no quote, another blockIndex is a string.
+- Observed result: `json: unknown field "$text"` from strict decoding. It never reaches the publishing sink.
+- This is a schema/model-output failure, not a source-store failure or fabricated success.
+
+Second tool call `call_H3g1poeMJTa4bBhzfBRu3S1r`:
+
+- Schema nesting/types repaired; same checkpoint ID is legal because the first attempt was rejected and not committed.
+- Twelve source links in the candidate were independently matched to host source IDs and source paragraphs. Eleven quotes are literal substrings. One review quote stitches two source portions with inserted `……`: “评审意见要求上线七月以留出接口联调时间……外部接口的负责人尚未确认。” The supplied paragraph contains intervening budget text and no such ellipsis, so this is not a literal quote.
+- Observed result: `checkpoint.source: quote not in provided block`. The rejection is correct. Details in `extra-a01-rejected-source-checks.json`.
+- Manager validates all candidate sources before committing any finding. This later invalid quote rejects the entire publication, including earlier valid records; no partial leak occurs. Neither call consumes an accepted-checkpoint count.
+
+## Persistent transactionality and truthful state
+
+Read the isolated SQLite DB with mode=ro and extracted exact card events to `extra-a01-persisted-events.json`:
+
+| Seq | Revision | State | Accepted checkpoints/findings |
+|---|---:|---|---|
+| 1 | 1 | queued | 0/0 |
+| 3 | 2 | running | 0/0 |
+| 418 | 3 | running activity | 0/0 |
+| 422 | 4 | running activity | 0/0 |
+| 426 | 5 | failed | 0/0 |
+| 429 | 6 | failed/reported | 0/0 |
+
+Registration05:09:53.281418215UTC; failed projection05:10:42.845780033UTC, approximately49.56s later. Rejected tool attempts only coexist with real activity updates; there is no findings publication in any persistent revision. Card outcomeCompleteness=none, executionState=failed, phase/focus state “执行已停止.” Exact run error is retained as code `run.subagent_failed`, message `budget_exceeded: run token budget exhausted`. Foreground remains completed. No malformed or stitched-quote finding was displayed as an accepted result.
+
+A truthful-copy edge remains on this sample's production SHA: conclusion says “已有发现已保留” and nextStep references existing findings even though there are none; remaining still says “待任务进展后更新” after execution stopped. Main state is honest, but these subordinate strings imply saved findings/future progression. Reported to executor for a host-owned zero-findings branch: state that no accepted phase result exists, clear obsolete process Remaining, and suggest inspecting materials/new request only when desired. Preserve this historical raw sample; do not rewrite it after correction.
+
+The material notice contains the correctly failed summary but retains generic deliverable-selection prompt text. Because the controlled notice model is disabled, no actual notification-model behavior is established by this sample. Do not claim it would correctly offer follow-up actions based on nonexistent results.
+
+## Visible answer quality
+
+There is no child final result or accepted checkpoint to grade as useful. The foreground provides factual file overviews and preserves missing approval/date/acceptance information, but repeats known unsupported interpretations: “分期vs一次性” implies the requirements demanded one-phase delivery; “剩余4项（含3项依赖接口）” assigns external dependencies to phase2 without evidence; “后续…需按会议版执行” chooses authority not supplied by the materials. It again duplicates the immediate answer after the stop-hook reminder. Questions later asking whether those assignments are correct do not remove earlier assertions.
+
+The semantic failures in rejected candidate text were not promoted to material records. Literal-source canonicalization is therefore not exercised by a successful actual publication in this extra sample. It remains mechanism-tested, not real-model demonstrated here.
+
+## Fixture-only code review and independent replay
+
+31efafa adds explicit authorizedA01 selection/output name, refuses to overwrite an existing completed sample file, and installs `limitExtraA01Runs` before the shared provider wrapper. The guard reads authoritative persisted run records in the isolated default-owner evaluation space and rejects a third run or query failure before forwarding a paid request. It deliberately does not rely on prompt fingerprints. The zero-token notice follows real work and makes no provider request.
+
+Independent REPLAY on exact31efafa before/after: `go test ./internal/server -tags live -run '^(TestExtraA01LimitsActualRunIDs|TestEvaluationPreservesProductionOutputCap)$' -count=1 -v` — exit0. Raw output `extra-a01-guard-replay.log`. These tests use fake streams/httptest, not paid models. The third-run guard and output-cap780 preservation both pass.
+
+## Incoming settings/SSE compatibility notes
+
+Read upstream Core0f94e4b and Backend741da4a changes without altering refs. Effective settings now merge persisted system scope and strip managed user/project overrides. Models and runtime.mock become system-managed; materialIntake, materialIntakeChars and context.maxAttachmentChars remain ordinary scoped values. U2 material hashing already consumes effective attachment cap. Current eval fixture's trusted model configuration must be seeded through system override/system scope after merge rather than now-rejected MergeUser. No route credential was read or printed.
+
+Merge must retain TaskSnapshot/DocumentOrigin while adding SettingDesc.Scope, and retain source visibility + subscribe-before-ready alongside upstream SSE session rechecks. These are compatibility observations, not a completed review/replay of the not-yet-merged composition. No broad audit of the upstream settings feature is claimed.
+
+## Verdict / declined scope
+
+**Extra A01: FAIL.** Zero accepted useful checkpoints, no child final answer, child ended on budget. Strict argument/source rejection, whole-publication consistency, run isolation/count and observed usage accounting behave as expected; those safety properties do not satisfy output quality acceptance. The two-run/60000 observed limit is met, not a reason for further retries.
+
+Overall U3 quality gate remains **not passed**. No more paid samples are authorized in this review. Declined: monetary billing (not provided); normal notification-model quality (suppressed); new merged component/runtime behavior (await exact composition); final platform CI/merge (executor gate). Retain the failure as concrete, reviewable evidence.
@@ -14,6 +14,7 @@ SELF原始命令输出由Dev的make evidence工具写入任务卡，不能升级

 | 对象及来源 | 检查 | 当前结论 |
 |---|---|---|
+| SELF，当前组合Devd06792c，Corebeafd75/Backende69865e/FE0bbccbe/UI3b34dfd/Docs1b296a1 | `make test && make build` | PASS：治理312、FE224、UI156，后端及构建通过；SELF原始收据`fast-slow-final-current-combination-serial.json`。当前模型质量仍FAIL。 |
 | SELF，最终候选Dev7fd5288，Core6e38574/Backend6a2d85c/FE3087b06/UI3b34dfd/Docs53a7d37 | make test && make build | PASS；治理311、FE224、UI156及完整后端/构建；收据fast-slow-final-fixed-combination.json |
 | SELF，Dev ed7f1c5，Core6e38574/Backendf6b5360/FE4d1d64a/UI3b34dfd/Docs59db26e | make test && make build | PASS；治理311，FE224，UI156；最后材料原文与评测上限修复是后续独立增量，不把旧SHA当新SHA |
 | SELF，Backend ee65604/2a64432增量 | go test ./...；live真实HTTP夹具780/5000/0输出限额测试 | PASS；完整后端、防止原文添加批准状态、停止后继及精确错误记录的回归通过 |
@@ -26,7 +27,7 @@ SELF原始命令输出由Dev的make evidence工具写入任务卡，不能升级

 独立报告原件：`/tmp/fast-slow-u2-u3-20261013/review/fix-review.md`、`model-assessment.md`、`model-assessment-remaining.md`。真实模型数据、执行事件和逐调用Usage均保留，凭据不在报告或源码中。

-代码复审归档：[独立复审记录](2026-10-13-fast-slow-u2-u3-code-review.md)。其结论是代码PASS、模型出口仍未通过，未把代码与模型测试混为一个完成状态。
+代码复审归档：[独立复审记录](2026-10-13-fast-slow-u2-u3-code-review.md)。新增模型样本的逐项独立判断见[额外A01评估](2026-10-13-fast-slow-u2-u3-extra-a01-review.md)。其结论是代码PASS、模型出口仍未通过，未把代码与模型测试混为一个完成状态。

 ## 原六样本的真实模型结果

@@ -56,3 +57,5 @@ SELF原始命令输出由Dev的make evidence工具写入任务卡，不能升级
 用户看到的Codex会话错误经当前chat元数据确认：上一轮回答流中途断开（request ID10e3777f-9b5e-4fc5-9dd2-132405d8c669）；当前chat恢复为active，源码及证据保留。该会话传输错误与应用模型验收失败分别记录。

 下一步建议先缩小单次检查点的复杂度，并依据实际输入开销调整每run预算分配，形成可评审的修订方案和有限复验预算；来源合法性保持严格，不能通过接受省略/拼接引用降低门禁。新的付费样本及原验收条件调整需明确授权。
+
+最新组合真实HTTP浏览器：Backende69865e/Corebeafd75/FE0bbccbe，两项来源预览和后台不抢焦点用例均PASS，SELF收据`fast-slow-browser-current-upstream.json`。首次组合检查遇到已自行消失的短暂Git索引锁，串行重跑后通过；未删除锁。