amos 9d633e9e67 Batch 15 (20261023-C-hwe-scale-15): 320 HWE-derived repair tasks
One user-space repair task per eligible HWE debug-diag / compile-deploy base
ID: 320 admitted (C 218, C++ 97, build scripts 5), 1 skipped by the content
filter. Each ships a public spec with a check-to-basis table, at least two
defects with single-fix fixtures, a symptom-masking fixture and an alternative
correct implementation. Claude re-checked all 320: structure, the HWE quality
items and trace6 Docker calibration receipts against current digests
(320/320 USABLE). Difficulty is unmeasured. Known follow-up: 212 tasks use 120
ad-hoc family names; the next batch maps them to the canonical list.
The batch's 1,256-line progress log stays in its factory report and the
36-local branch batch15-hwe.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-27 01:55:56 +08:00

7.0 KiB
Raw Permalink Blame History

concur-level-stack author specification

Status: CALIBRATED / UNMEASURED. Route C, original C++17 software replacement of hwe/cc-lo-013. The HWE material supplies only the lower-to-higher lock topic. This package is not a recovered production incident. The controlled port records exclusive owners without scheduling, ARM hardware, real mutex timing, or real thread interleavings.

Behavior and evidence basis

Grader behavior Public basis
Owner and lock domains; six unequal levels; exclusive port ownership template/docs/contract.md:3
Lower-to-higher acquisition per owner, independent owner interleaving template/docs/contract.md:4
LIFO release and actual remaining stack after nonconsecutive levels template/docs/contract.md:5
Batch atomicity, first failing step template/docs/contract.md:6, template/docs/contract.md:9
Release all, reset, no change to other owners template/docs/contract.md:7
Return codes and failed-state preservation template/docs/contract.md:8, template/docs/contract.md:9
View, port owner, count, snapshot template/docs/contract.md:10; template/docs/protocol.md:5-6
Audit, acquire/release previews, counters template/docs/contract.md:11; template/docs/protocol.md:9-12
One stdout observation per command, stderr allowed, exit status template/docs/protocol.md:3-8
Build and submitted files template/README.md:5 and prompt.md:5

The I audit does not replace independent state comparison: the Python model computes all expected statuses, stack contents and port slots. The private C++ console only prints raw observations; no hidden expectation or assertion is compiled into it. The grader uses fixed-seed mixed traffic and compares complete numeric stdout fields, permits documented signs/leading zeros and horizontal separators, and ignores arbitrary stderr bytes. It separately checks real process exit, timeout, and cleanup status.

Template defects

  1. Cross-owner level leak. The template gates all owners against one current_level_. Trigger: owner 1 holds level 50 while owner 2 takes level 20; this is a documented independent owner case (contract.md:4-5). only_fix_restore repairs release restoration but leaves this defect.
  2. Guessed restoration. The template restores released_level - 1, not the top of the actual held stack. Trigger: take levels 10 and 50, drop 50, then take 20; all are legal (contract.md:3-5). only_fix_global removes the cross-owner leak but retains this defect.

No negative case relies on assertion failure, deadlock or a build error. All fixtures compile. mask_symptom alters one public example path without recording ownership and fails the full state observation. noop refuses core acquisitions. cheat reports success without owning a port lock. cheat_read_grader behaves as the gold implementation only when /grader/check.py is readable to the candidate; otherwise it refuses acquisitions. The fixed image must deny that read. valid_preflight scans physical port ownership and preflights an entire batch before mutation, unlike gold's per-owner cached levels and incremental rollback.

Fixture matrix

Fixture Expected Mechanism
gold PASS Per-owner level cache restored from actual held stack, transactional rollback
template FAIL Both defects present
noop FAIL Core acquisition returns busy
cheat FAIL Reports success without port ownership
cheat_read_grader FAIL Private grader read gate; candidate identity cannot read it
only_fix_global FAIL Guessed restoration remains
only_fix_restore FAIL Cross-owner shared level remains
mask_symptom FAIL Public case special branch leaves port unchanged
valid_preflight PASS Port scan and preflight batch, a different data/control strategy

Execution and admission

The grader rebuilds from candidate files plus frozen public support in a unique scratch. C++ is run only as a child of a candidate-identity supervisor. A Python model supplies the expected values. The supervisor carries no expectation, reports raw stdout/stderr lengths and hashes, authenticates PID/start identity, enforces deadlines, and reaps its child tree. Cleanup and independent strict lstat absence run under the same candidate identity in finally. All four phases must report complete, consistent metadata with zero exceptions; only a normal 0 compile/candidate/cleanup/absence result and all matching observations can PASS. Trusted local mode is only a trusted author check; fixed-image calibration is the isolation evidence.

The package remains UNMEASURED until real Agent samples are run and reviewed. The root will do the independent read-only review before deciding whether to use it for pilot sampling.

Actual calibration and diagnostic evidence

The final task digest is 41dd0bf196753f37a615aa0e191c417869cada9a761bd84c4a4fa95271c85de5. The trusted-local receipt is runs/20261023-C-hwe-scale-15/cc-lo-013/local-final/calibration-44c30692ced140eb8f1b72dc32507bbc/receipt.json; the fixed-image receipt is runs/20261023-C-hwe-scale-15/cc-lo-013/docker-final/calibration-5e44e8714814448b9dd3655e4ba4d821/receipt.json. Each is PASS with 45/45 trials matching fixture expectations and cleanup confirmed. Gold and valid_preflight each passed five times per gate; all seven negative classes failed five times per gate. All 90 compact metadata logs decoded with complete compile, candidate, cleanup, and independent absence phases. The image ID is sha256:1519f8daa403998e5401eeb70b973e8aac8346e25a937033bc502dc5de1641a1.

Fixed-image core._grade diagnostics used 18 complete loadable task copies under the author directory. Gold, arbitrary 0–255 stderr, 17 MiB stderr, and a closed-stdio setsid descendant passed. Compiler FIFO, candidate-held pipe and stdin backpressure reached actual production deadlines and failed after child cleanup; cleanup nonzero, cleanup timeout, cleanup no-op with absence failure, absence timeout, preparation fault, empty OSError after success/cleanup/absence, missing and mistyped wire fields, and supervisor nonzero all failed. Each copy and attempt retains actual package, check/context/candidate hashes, raw wrapper log hashes, stage metadata and deadlines in factory/20261023-C-hwe-scale-15/subagents/cc-lo-013/diagnostics/summary.json. Synthetic process-graph tests cover zombie ancestors, reused PID/start tokens, and fresh signal authorization. No real PID exhaustion or unrelated process was touched.

The first local receipt at the old digest was PASS; the first Docker receipt at the old digest was FAIL because scratch preparation called nonexistent Path.chown. Both are preserved as excluded history. The checker uses os.chown in the final identity and passed the fixed-image diagnostic and formal gate. Full candidate stdout/stderr bytes are consumed for checking and bound by length/SHA metadata, but are not archived as standalone raw files; the wrapper's stdout/stderr files are retained. No real Agent, ARM, or threading-timing claim is made.