amos 4f17b2afe5 Batch 20 (20261027-C-family-20): map batch 15's ad-hoc families to canonical ones
210 batch-15 tasks used 119 ad-hoc family names. Each now carries one of the
canonical root-cause families (11 of the 19 are used), with version + 1 and a
spec.md note, and was recalibrated through both gates. Claude checked that only
family, version and spec.md changed, that every family is canonical, and the
new trace6 Docker receipts against current digests (210/210 USABLE).
Squashed from the 36-local branch batch20-family.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-27 05:58:17 +08:00

9.0 KiB
Raw Permalink Blame History

state-dispatch-effects author specification

1. Identity and source

Version 1, family state-dispatch, language C11, source hwe/dbg-gdb-short-circuit, source/split group dbg-gdb-short-circuit. Route C original, authorized HWE theme material; no historical incident claim. The source prose asks for a required increment while its q=5 expectation retains the skipped increment. Neither contradictory output is copied. This task uses only the coherent knowledge point: conditional evaluation must not suppress required state effects. No source scripts were executed, no external URL or new dependency is needed.

2. Feasibility and scope

A deterministic four-lane event dispatcher models message queues, leases, flow control and a receiver callback. Software substitution retains both conditional-evaluation defects without a board, physical timing, kernel state, network or memory corruption. All queues, receiver refusal decisions and resets are controlled and observable. This is a software reliability exercise and does not prove ARM hardware or debugger operation. The original source's CMake requirement is background; this original project uses the fixed runtime's gcc/make.

The public engineering scope includes ring queues, lease compaction, cancellation of repeated ids, transactional queue handoff, validated commands, delivery policies and receiver accounting. Only four submitted files are frozen. Supplied driver/receiver support is frozen verbatim from public files in context.json. The driver prints raw observations and contains no expected values or assertions. Private Python computes expected observations from the contract without running gold.

3. Initial defects and independence

  1. Maintenance: tick uses lane->enabled && maintain_lane(lane), so disabled lanes skip the compulsory pulse and lease aging. Trigger: disable a lane, publish a lease-bearing message, tick. This is visible in demo.txt. Basis C4, contract.md:13. only_fix_maintenance fixes this defect but remains wrong in broadcast.
  2. Broadcast: mode 1 uses any = any || offer_lane(...), skipping required later offers after one succeeds. Trigger: set nondefault mode 1, publish to multiple ready lanes, tick. Basis C5, contract.md:16. only_fix_broadcast fixes this defect but remains wrong for disabled-lane maintenance.

They involve separate phases and state effects. Both only-fix samples build and fail through normal-exit observations. No compile/link failures are used as negative examples.

4. Public criterion matrix

All paths below are relative to template/. Hidden inputs vary combinations, not expected behavior. The oracle compares only live queue entries and integer tokens; unused storage and stderr diagnostics are unrestricted.

Evaluated behavior Exact public basis
Four lanes, queue capacity 8, startup/reset defaults docs/contract.md:4, dispatch.h
id/value/lease limits and full command stream bound docs/contract.md:4
Blank/comment lines produce no observation docs/contract.md:7
Decimal signs, surrounding whitespace, inline comments, EOF without newline docs/contract.md:7
Malformed integer, 32-bit parser overflow, trailing tokens and unknown commands rejected docs/contract.md:7; C11 limits.h used by public codec.c
Invalid command increments only invalid; argument/arity validation before mutation docs/contract.md:7, docs/contract.md:21
Mode 0/1, lane enable, credit 0..8, period 0..8 configuration docs/contract.md:10
L preserves pending messages and all existing counters docs/contract.md:10
Credit addition overflow returns -3 without mutation docs/contract.md:10
Discard clears lane and counts cancelled messages docs/contract.md:10
Mask publication duplicates in selected FIFOs, repeated ids allowed docs/contract.md:10
Publication capacity refusal is atomic and increments blocked exactly once docs/contract.md:10
Cancellation removes every selected matching id, preserves other order docs/contract.md:10
Transfer preserves FIFO/lease; insufficient source -4 precedes destination -2 docs/contract.md:10
Transfer refusal atomicity and successful moved_in/moved_out docs/contract.md:10
Business refusals do not increment invalid docs/contract.md:10
Every tick pulses and ages all lanes including disabled/no-credit docs/contract.md:13
All expiration precedes delivery, exact lease-one boundary, stable survivors docs/contract.md:13
Cyclic lane ordering and at most one offer per lane docs/contract.md:16
Unicast stops after first success; rejected offers continue scan docs/contract.md:16
Broadcast continues through all lanes after success docs/contract.md:16
Accepted pop/credit/delivered and rejected retained queue/credit/refused docs/contract.md:16
Cursor advances unconditionally once per tick docs/contract.md:16
Port attempt/accepted/rejected and latest offered id/value docs/contract.md:18
Rejection period and signed sum/modular digest docs/contract.md:18
Public C11 API, build, FIFO order after wrap/compaction docs/contract.md:21, Makefile
Complete S/L/Q/V observation fields and live-range-only comparison docs/contract.md:23
Token whitespace flexibility, stderr allowed, reset snapshot docs/contract.md:23

5. Samples and mechanism

All fixture directories contain exactly dispatch.c, queue.c, codec.c, dispatch.h. Gold separates maintenance from eligibility and always performs each broadcast offer before aggregating its result. Template has both defects. Noop returns success without applying commands. Cheat special-cases lane zero, masking the default demo while retaining wrong broadcast/other-lane behavior. The read-grader sample runs gold only when it can open the protected check; otherwise it exits 71 with grader unreadable. Trusted local access is the documented exception; isolated Docker must reject it. The two only-fix samples each preserve the other defect. Mask-symptom deletes disabled queues rather than aging individual leases and recording maintenance. Valid-worklist uses contiguous FIFO arrays with shifting deletion and a prepared lane worklist, rather than gold's wrapped ring and direct scanning; it also emits permitted stderr diagnostics.

6. Verification and lifecycle

The private standard-library Python model generates 10000 observed commands with fixed seed 884920, including directed independent defects, empty/full queues, atomic refusals, repeated ids, transfer/cancellation, ring wrap, every mode, negative/positive extremes, periods, malformed commands, reset and long histories. The public C driver and receiver only expose raw observations; gold is never used for expected values. All material and stdin are prepared before compilation. Compiler and program run through argv subprocesses; after the first child begins the parent never reads/writes/traverses candidate-controlled scratch. A finally block always runs an argv Python cleanup child under the same candidate identity and checks rc; cleanup failure cannot PASS. Scratch paths, compile rc, candidate rc, controlled refusal and first normal-exit difference are distinct diagnostics.

verify.py is byte-identical to rollout/bos_verify.py; grader_files contains only check.py and context.json. Docker UID65534 isolation and unreadable grader are required. Timeout margins are 30s compile, 20s candidate, 10s cleanup; full verifier 100s/manifest 120s. Gate evidence uses each of nine classes five times in the formal local and fixed-image calibration entries. The two gates bind the current task digest. Only the five trusted-local read-grader PASS trials are allowed exceptions.

7. Status and limits

Status CALIBRATED/UNMEASURED: current fixed-image gate passed 45/45; trusted-local gate has only its five documented read-grader PASS exceptions; never ADMITTED. Real Agent sampling, native rollout evidence, hardware, network and actual debugger use are unmeasured. The harness exercises APIs through the public command driver; direct misuse of internal queue helpers and NULL pointer callers are outside this task's tested contract. The stream contains exactly 10000 observed commands and no concurrent operations. Root performs independent review and the shared full-suite checkpoint; this author does not modify the total ledger, run model sampling or commit.

Final evidence: local /home/ming01.wang/embedded-rollout/.tmp/wt-batch15/runs/20261023-C-hwe-scale-15/dbg-gdb-short-circuit/local/calibration-fdafbcf5aa74474b95265b5b40c3983c/receipt.json; Docker /home/ming01.wang/embedded-rollout/.tmp/wt-batch15/runs/20261023-C-hwe-scale-15/dbg-gdb-short-circuit/docker/calibration-b5368600c80d4b12ac99ee7df8365fd6/receipt.json. Both bind task digest b16cd942470e1ef9a03eac4ef76017343c60f8fd61d412ad8e9a3340899dbd10. Candidate-identity cleanup rc=0 in all 90 formal trials; exact 49 logged Docker objects and 54 logged host scratch paths freshly absent. Author review complete; independent root review pending. No full-suite or model sampling was run by this author. 族名规范化:state-dispatch → state-machine-timing(2026-10-27)