100 tasks from the HWE arm-linux code-gen pool (outside the lk/man/pth range given to batch 22), each with at least one independent contract-edge defect and plain public docs. Claude reviewed 100/100 USABLE against current digests; no ID or source overlap with main or batch 22. Difficulty is unmeasured. progress.md records the wrap-up: all Codex sessions and Qwen schedulers stopped for the user's rollout, Qwen results so far (57 HARD of 373), and what is left on each batch branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
17 KiB
Task specification: dev-deferred-callback-gates
1. Identity and source
- Version 1; family device-io; source exactly
hwe/ldd3-tasklet-disable-enable. - Route C, original software task authored from HWE public/hard, non-held-out source direction. The source kernel exercise and reference file are read-only local materials; none of their scripts, checks or driver source were executed or incorporated.
- Source group and split key: ldd3-tasklet-disable-enable. Root-cause family: deferred callback batch retirement and paired gate depth. No claim of reproducing a real incident. No private user conversation or secrets included.
- Software-only substitution: sequential explicit logical steps replace tasklet execution. Pending coalescing, disable/enable, already-started work and teardown boundaries remain observable. No kernel module, device, timing, board or privilege is needed.
- Author: assigned single-task Codex agent. Independent review: pending root review. Status: CALIBRATED, independent root round 1 review pending. Difficulty UNMEASURED; no Qwen, real Agent sampling or hardware validation.
The synthetic incident is loss of accepted work arriving while an earlier callback is active. A second independent fault breaks exact nested-disable pairing at depth 2 to 1; this boundary is specified but absent from smoke. Both faults are cross-cutting between gate/registration state, dispatch and independent observations.
2. Initial state and environment
The public workspace contains the five candidate files, README.md, docs/contract.md, Makefile, observer.cpp and smoke.in. The observer only forwards raw stdin arguments to public APIs and prints raw return records; it contains no expected values, generated scenarios, assertions or verdicts. The candidate files are deferred.hpp, engine_internal.hpp, registry.cpp, dispatch.cpp, inspect.cpp.
C++17, ordinary Linux user space, standard library, g++ offline build. No external services or hardware. Formal image is trace6 ccr-29eug8s3-pub.cnc.bj.baidubce.com/algo/bench-runtime@sha256:1519f8daa403998e5401eeb70b973e8aac8346e25a937033bc502dc5de1641a1. Actual image ID and task/verifier/manifest digests are bound by calibration receipts, not inferred from the tag. Grade limit 120 seconds; compiler 30 seconds; each candidate process 10 seconds. Docker grade resource policy is platform-managed (512 MiB, 1 CPU).
Protected assets are check.py, verify.py, gold and all fixed alternate/negative solutions. None enters template. check.py generates fixed-seed inputs, routes selected-instance commands to two independent per-request Python reference models, computes expected results, and compiles a byte-identical copy of the public observer with candidate sources. It reads/writes all scratch contents before the first compiler process; after that it only compares captured bytes in memory. Scratch uses tempfile.mkdtemp(dir=Path.cwd()) under the wrapper's exclusive root, and wrapper teardown removes the tree. No root candidate import, dynamic loading, shell invocation, or post-run scratch access occurs.
3. Public requirement and reproduction
Prompt only reports the observable missing-delivery symptom, points to make && ./deferred-observer < smoke.in, requires every clause of docs/contract.md, and lists the five delivery files. It does not name the faulty method, implementation file or nested boundary. Smoke contains no disable/enable commands. Template's first two requests are accepted; the first completes; the next step incorrectly reports IDLE. Correct final smoke steps include a second START with sum 13/value 20.
Build argv: ["g++", "-std=c++17", "-O2", "-Wall", "-Wextra", "-pedantic", "observer.cpp", "registry.cpp", "dispatch.cpp", "inspect.cpp", "-o", "deferred-observer"]. Normal compile exit 0. Run argv ["./deferred-observer"], stdin well-formed decimal command lines. Every command produces exactly one decimal observation line; candidate stderr must be empty and the process exits 0 even for rejected API calls. Malformed lexical inputs and out-of-int32 int arguments are explicitly outside the protocol; out-of-domain values within their C++ types are inside it.
4. Check-to-evidence matrix
All cited paths are public workspace paths under template/. Every comparison is exact, with no numeric tolerance. Every method is called directly by the observer; no snapshot substitutes for another method's return.
| Check class | Public evidence | Protected scenario / method | Observation |
|---|---|---|---|
| Complete declared API, fresh engine | deferred.hpp:76; docs/contract.md:7 | interleaved_engine_independence keeps two objects alive in the same process; every command type | actual API return records |
| Slot endpoints, lifecycle and generation | docs/contract.md:9 | domains_precedence_atomicity; lifecycle_priority_nosync_kill | IDs 0,15 accepted; -1,16/int32 extremes rejected; reinstall generation |
| Numeric statuses and actions | docs/contract.md:11 | all ten scenarios | numeric codes, START/FINISH/IDLE |
| Install priority domain endpoints | docs/contract.md:15 | domains_precedence_atomicity | 0,3 valid; -1,4/int32 extremes invalid |
| Install quota domain endpoints | docs/contract.md:15 | quota_signed_ledger; domains_precedence_atomicity | 1,32 valid; 0,33 invalid |
| Cookie domain and signed endpoints | docs/contract.md:15 | domains_precedence_atomicity; maximum_stream_wide_totals | ±1000000 accepted, ±1000001/int64 extremes rejected |
| Install precedence, EXISTS, atomic reset | docs/contract.md:15 | domains_precedence_atomicity; lifecycle_priority_nosync_kill | ID→numeric→existing; ledger/config/generation reset |
| Post numeric and lifecycle precedence | docs/contract.md:17 | domains_precedence_atomicity | BAD_ID→BAD_VALUE→MISSING/DEAD_SLOT→LIMIT |
| Pending count/sum/ticket coalescing | docs/contract.md:17 | pending_witness; quota_signed_ledger; fifo_and_fresh_engine | exact sums/counts, retained ticket, fresh-ticket monotonicity |
| Active batch does not consume pending quota | docs/contract.md:17 | quota_signed_ledger; fifo_and_fresh_engine | quota 1 and 32 while active, exact overflow rejection |
| Error atomicity and mutator error count | docs/contract.md:19 | observe() after rejected calls | direct G/V/H; exactly one errors increment, no ticket/event changes |
| Query rejection never counts as mutation | docs/contract.md:19 | domains_precedence_atomicity; observe() | repeated V/H failures followed by direct G |
| Synchronous disable BUSY before depth LIMIT | docs/contract.md:23 | domains_precedence_atomicity | active at depth4 returns BUSY, N returns LIMIT |
| nosync never stops already-started callback | docs/contract.md:23 | lifecycle_priority_nosync_kill; quota_signed_ledger | active FINISH occurs at nonzero depth; pending remains blocked |
| Exact nested pairing and zero enable | docs/contract.md:25 | depth_pairing_witness | depth2→1 IDLE preserving merged work; depth1→0 START; depth0 UNBALANCED |
| Completion preserves newer pending | docs/contract.md:29 | pending_witness; quota_signed_ledger | FINISH original count/sum; next START pending batch |
| Single active worker and complete-before-start | docs/contract.md:29 | lifecycle_priority_nosync_kill; seeded_combinations | direct Action and running fields |
| Priority ordering, FIFO and disabled skip | docs/contract.md:31 | lifecycle_priority_nosync_kill; fifo_and_fresh_engine | selected IDs, tickets and batch movement |
| No eligible work IDLE without event | docs/contract.md:31 | initial S, depth_pairing_witness, long-run tail | exact idle record plus unchanged event/ticket counters |
| Per-generation transformed value and wide totals | docs/contract.md:33 | quota_signed_ledger; maximum_stream_wide_totals | sum+cookie*count, totals 7040000000 beyond signed32 |
| Kill active finish-before-kill | docs/contract.md:37 | lifecycle_priority_nosync_kill | completion ledger and ordered FINISH/KILL events |
| Kill pending cancellation, other active unaffected | docs/contract.md:37 | lifecycle_priority_nosync_kill | DEAD zero work/depth; other active still finishes |
| Repeated kill and reinstall do not revive work | docs/contract.md:37 | lifecycle_priority_nosync_kill; event_ring_reuse | DEAD_SLOT and clean generation reuse |
| Direct view including dead ledger | docs/contract.md:41 | observe(), final per-slot views | all 16 View fields; nonzero sentinel unchanged on error |
| Direct summary counters and query purity | docs/contract.md:43 | observe(), maximum_stream_wide_totals | all 10 Summary fields, no inferred substitute |
| Event ring capacity, order, serial and error outputs | docs/contract.md:45 | event_ring_reuse; H0/H15 after transitions | all 7 fields; 16 retained, oldest drop; -1/16 reject; missing index sentinel |
| Well-formed transport, typed integer bounds | docs/contract.md:49 | domains_precedence_atomicity; all cases | raw parsing preserves full int64 values and all arguments |
| Exact stdout bytes, empty stderr, normal exit | docs/contract.md:51 | every run; cheat_stderr | byte-for-byte stdout including final LF, candidate stderr empty, run0 |
| Same-process receiver selection and per-instance state | docs/contract.md:53; README.md:9 | interleaved_engine_independence, U0/U1 at both selector domain endpoints | fresh second instance, interleaved pending/active/gates/errors/lifecycle/tickets and event serials remain independent |
| Maximum 4096-command stream | docs/contract.md:9; docs/contract.md:49 | maximum_stream_wide_totals | 4096 calls, no cutoff or integer truncation |
The final tie-breaker ID is documented but equal pending tickets cannot be produced for distinct slots through this API; all reachable sorting cases are covered. Compiler flags do not require -Werror. Standard-compliant alternate implementations are accepted without source inspection or code-shape assertions.
5. Fixed solutions and calibration
| Fixture | Expected | Role and concrete witness |
|---|---|---|
| gold | PASS | In-place per-slot pending/active aggregates, completion ledger, ring |
| template | FAIL | Both independent bugs; pending_witness first rejects |
| noop | FAIL | Accepts posts without retaining requests; first START is IDLE |
| cheat | FAIL | Hardcodes the public first START record; required FINISH differs |
| cheat_read_grader | FAIL (Docker) | Uses correct behavior only after reading grader; denied grader leads exit23 and grader unreadable |
| fix_pending | FAIL | Fixes only completion clearing; depth_pairing_witness rejects depth2→1 |
| fix_depth | FAIL | Fixes only exact gate subtraction; pending_witness still loses newer work |
| mask_symptom | FAIL | Folds new request into executing batch; wrong FINISH count/sum/value |
| valid_request_log | PASS | Authoritative individual Request records with waiting/executing/delivered/canceled phases and gate-token stack; batch values and completed ledgers computed by grouping requests; no gold pending/active aggregate authority, no incrementally maintained delivery totals, and sort-based arbitration |
| cheat_stderr | FAIL | Correct numerical behavior but emits only forbidden candidate stderr diagnostic |
| cheat_shared_instance | FAIL | Shares a singleton Impl across both Engine objects; all single-instance cases pass, but U1 exposes instance0 state |
Every fixed solution contains all five candidate files and compiles/links against the same raw observer. Ordinary negatives complete every scenario with run0 and fail by observations. The read-grader negative alone may terminate after denial. On trusted Linux host it can read the grader and PASS; preserve this specified exception and overall host mismatch rather than alter expected values or receipts. Docker must match all 55 trials.
Each backend calibration uses five fresh resets per fixture (55 trials). Final formal gate receipts and real exit codes are appended below. First preflight and coverage preflight are immutable records under factory/20261026-C-hwe-codegen-18/subagents/ldd3-tasklet-disable-enable/. Final task digest covers the public docs, private check and all solutions.
6. Sampling and limits
No Qwen/GLM or any real Agent is run in this assignment; difficulty UNMEASURED. No native model/tool trace or hardware claim. No Salty upload/download or cold migration is performed by this author. No credentials, old task edits, platform changes, all-repository tests, Git commit or push. Wrapper scratch is cleaned by wrapper teardown; calibration containers are managed and removed by platform and checked only for this task's known names.
7. Review and admission
Author preflight: gold and alternate pass; both single-fix variants fail their independent remaining witness; all ordinary negative build0/run0 checks pass and observations reject them. Final coverage review is recorded in coverage-audit.json before dual gates. Independent review remains with root. Task is CALIBRATED only after matching Docker receipts and not ADMITTED solely from software calibration. Known restriction: correctness is only for well-formed raw protocol and explicit logical model; no real tasklet concurrency or hardware.
Root review round 1 rework
Root found that launching a fresh process for every scenario did not test the declared same-process Engine independence. The public observer and R7/README now expose U 0/U 1, with exactly 0 or 1 guaranteed as selector-domain inputs; the adapter echoes the selected index, forwards APIs to it, and preserves both live objects. No new candidate API or candidate algorithm change is required. The public observer and checker literal are synchronized.
interleaved_engine_independence first creates instance0 pending/active work, disabled gates, counters, tickets and events; selects instance1 and reads empty initial state; performs independent errors, posting and active work; repeatedly returns to instance0; kills/reinstalls instance1 and proves neither its lifecycle, counters, ticket allocation nor event serial affects instance0. It then advances both workers separately. A dedicated complete-build/run0 negative cheat_shared_instance passes every previous single-instance scenario and fails only this new witness. The complete matrix now has 30 rows; all scenarios remain within 4096 commands.
Round0 receipts and source checker/spec/author artifacts are retained unchanged in the original runs and factory root-review-round0/. The replacement final gates below bind the complete coverage freeze.
Final replacement calibration evidence: root review round 1 (2026-10-27)
- Current digest:
3a61d2a21fb3c07b8c05efcb90ba67c198618bbf181cbe20267f9248679dc850. Both replacement backend receipts match the current task package. The previous digest and its original 100 trials remain unchanged at their original run paths; author artifacts and previous spec/checker are also preserved in factory root-review-round0/. - Candidate sources remain exactly the same for all ten previous fixtures, verified by per-fixture candidate digests against the old receipt. Candidate effective source stays 573 nonblank/noncomment C++ lines over five files. Unique project C++ source including the expanded raw observer is 627 lines. Fixture copies are excluded.
- The checker now runs 10 scenarios including
interleaved_engine_independence(225 commands). Both receiver selectors 0 and 1 and all instance independence claims are directly observed in one process. The matrix has 30 actual evidence rows with public line references.cheat_shared_instancematches the prior nine scenarios and fails only this added witness, after build0 and normal run0. - Host receipt:
runs/20261026-C-hwe-codegen-18/ldd3-tasklet-disable-enable/local-round1-20261027T063209-229f8d3c/calibration-427ba4f881944424982086f58bfc1603/receipt.json. 55 trials, real CLI exit 1, overall FAIL, cleanup confirmed. 50/55 match; only the five allowed trusted-host read-grader exceptions PASS unexpectedly. Gold/valid 5/5 PASS, ordinary negatives 5/5 FAIL; no expected values or original receipts were modified. - Docker receipt:
runs/20261026-C-hwe-codegen-18/ldd3-tasklet-disable-enable/docker-round1-20261027T063209-229f8d3c/calibration-d130bd17bdba465e815c2aa06bf5cd9f/receipt.json. 55/55 trials match, real CLI exit 0, overall PASS, cleanup confirmed, imagesha256:1519f8daa403998e5401eeb70b973e8aac8346e25a937033bc502dc5de1641a1. Gold/valid each 5 PASS, nine negatives each 5 FAIL. - All 110 current raw stdout/stderr pairs match their receipts. All candidates compile with exit0. Every ordinary fixture executes all ten scenarios with run0. Only the explicitly allowed Docker read-grader negative exits23 after denial. Stderr-only rejection remains explicit.
- All 55 exact current-round container names were inspected and absent; no owned network was created. Current cleanup.json points to uniquely named raw inspect outputs. The previous 50-container cleanup evidence and raw output files remain preserved; this check claims no global resource discovery.
- State/report/coverage/author-review/trial/gate/cleanup artifacts have been updated for root round 1. No additional Agent, Qwen, all-repository test, shared code change, Git action or hardware execution occurred. Independent root review is pending; difficulty remains UNMEASURED.