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

6.7 KiB

resource-render-buffer

Status: CALIBRATED. Difficulty: UNMEASURED. Language: C11. Source route C original; source hwe/dbg-valgrind-leak-format-buffer; split group/base ID dbg-valgrind-leak-format-buffer. HWE only supplied the dynamic formatting-buffer ownership direction; this is a newly written synthetic engineering problem, not a replay or real incident. It does not claim valgrind, hardware, production isolation or a real Agent trace.

A bounded private-static fixed buffer port replaces malloc/free and a transmitter. The candidate still writes actual serialized content to its allocated port buffer and publishes it while live. The independently compiled support sees actual live flags, buffer bytes and publish copies; it never reads candidate counters. No real invalid heap access or buffer overrun is needed. C11, gcc and make suffice; the pinned image contains gcc/make but not CMake. All three candidate sources are compiled in the actual fixed command; public Makefile is frozen support, not a candidate.

Independent defects: (1) the success path leaves its allocation live, exposed by default make demo; only_fix_release repairs it. (2) a requested capacity equal to canonical text length is accepted even though it lacks NUL space; the bounded writer safely produces an incomplete serialization. This is outside the actual default replay and is explicitly an independent optional make capacity; only_fix_capacity repairs it. Failure paths and unused storage remain bounded. The optional replay has the exact-boundary case before a sufficient-capacity retry.

The Python oracle independently assembles canonical bytes and computes unsigned polynomial digest, accepted/rejected state and expected fixed-port observations. It uses no gold execution and has no candidate source interpretation. The public raw observer prints only raw candidate results and fixed-port observations. The deterministic suite contains 79 physical lines, covering default, required-1/equal/+1 capacities, zero and max capacity, all supported failures, binary/control/quote labels, int32 bounds, long leading zeros, lexical rejection, sequence ordering and long sessions.

Criterion Exact public basis
Physical line framing, horizontal separators, integer lexical domain and signed range template/docs/render-contract.md:3
Label hex decode and binary bytes, empty and maximum label template/docs/render-contract.md:10
Atomic invalid lines and no extra operations template/docs/render-contract.md:14
Sequence rejection, accepted/failed/successful/last counters template/docs/render-contract.md:16
Cumulative polynomial session digest over complete successful records template/docs/render-contract.md:20
Canonical field order and minimal numeric rendering template/docs/render-contract.md:23
Quote/backslash/lowercase unicode escapes and embedded NUL label template/docs/render-contract.md:26
Required length excludes NUL and exact capacity requires TRUNC template/docs/render-contract.md:28 and template/docs/render-contract.md:33
One allocation attempt, zero/forced failure precedence template/docs/render-contract.md:30
No publication on truncation; emitted0 template/docs/render-contract.md:34
Live complete publish, copied output bytes, sink failure and emitted count template/docs/render-contract.md:37
Exactly one release on every allocated path and no invalid port calls template/docs/render-contract.md:41
Actual fixed port counters/live state/digest and per-line output clear template/docs/render-contract.md:47
Raw observation row/field schema, numeric equivalent display, defined output bytes only template/docs/render-contract.md:58
Arbitrary stderr bytes accepted and process exit0 template/docs/render-contract.md:66
Default vs independently optional exact capacity scenario template/docs/render-contract.md:70

Fixture mechanisms: gold uses a counting/emitting writer followed by release; valid_two_pass constructs an escaped-label scratch representation and whole text with snprintf, then copies into a port allocation under a nested single cleanup block. Their algorithms, temporary representation and control flow differ. template retains both defects; only_fix_release and only_fix_capacity leave the independent other defect; noop returns an empty successful result without port behavior; cheat publishes a fixed default string; mask_symptom allocates the maximum capacity regardless of the requested limit; cheat_read_grader needs readable private check.py before behaving as gold and exits 71 otherwise. All sample directories contain only candidate_files.

Execution uses the root-reviewed same-task supplementary execution-only reference, extended with strictly None adopted wait start identities. Candidate-identity subreaper deadlines cover compiler20s, candidate8s, cleanup5s and independent strict-lstat absence5s, plus 3s recovery per stage. Parent validates complete typed identity/return/deadline/IO metadata and requires empty exception lists. It never touches the candidate tree after the first subprocess. All actual wire/IO lengths and digests bind captured bytes; compact marker stores complete metadata but no wire/IO body. Original inner wire/IO is not retained by formal grades; no claim of original-body preservation is made. Diagnostic actual packages and raw outer logs are retained separately.

Actual evidence: final fixed-image gate 45/45 expected PASS, cleanup confirmed; trusted-local has exactly the five allowed reader PASS exceptions (original overall FAIL retained), all other 40 match. Formal 360 stages comprise 355 leader0 and 5 Docker reader71. Nine-fixture preflight has compiler0 for every sample, leader0 for eight and reader71. Five fixed-image make builds and ten public replays preserve raw IO; only_fix_release passes default but fails optional exact capacity. Production matrix has 22 expected diagnostic results plus one retained/excluded diagnostic constructor SyntaxError, corrected in a single failed-case-only retry. Normal leader0/closed-stdio descendant, all-byte/17MiB stderr and long-zero stdout PASS; actual deadline/cleanup/strict-lstat/wire/exception failures reject. Pure helper proof has 74 cases bound to the current check. Precise cleanup inspection found all 83 recorded container names absent and no live matching local process tokens. Full report, exact package/context/raw bindings and honest original-inner-IO limitations are in the same-author report.json. Independent root read-only review remains pending. No model/Pi/Qwen sampling, hardware, all-repository suite, platform edit or git operation is authorized for this author. Calibration alone is CALIBRATED/UNMEASURED, never ADMITTED.