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

9.1 KiB

resource-line-buffer v1

Route C original; source and split group hwe/gi-leak-getline-loop. The source material supplied only the reusable getline ownership direction. Root reviewed the six source files and cold/local SHA before assignment (root-next-medium-source-proof-05.json); no source script was executed. This newly authored service, protocol, oracle and fixtures do not claim a real incident. Author: sole gi-leak-getline-loop worker. Independent review: root pending.

The deterministic C11 byte-source callback replaces a real stream; an observable allocator replaces external leak instrumentation. Dynamic line storage is still retained and reused across reads, so the same terminal ownership defect survives substitution. No board, ARM ABI, real file, GNU extension, network, clocks or asynchronous signal safety is assessed. Intended medium difficulty is UNMEASURED: no Agent sampling has been done. Status after gates is CALIBRATED/UNMEASURED, never ADMITTED.

The complete public workspace has four submitted modules, a public header, a raw observation program, Makefile and docs/contract.md. Files are not imported from another task. The fixed image provides gcc 14.2.0 and GNU Make 4.4.1, verified by an explicit project-named read-only tool probe; exact command and original output are in the author's tool-probe directory. Local public make clean / make / make demo have original stdout/stderr in public-commands.json and public-*.stdout/stderr. The demo exposes a live allocation after a text job.

The first defect is in the job's terminal ownership branch: normal EOF or stop returns without destroying a nonempty line allocation. only_fix_release repairs that branch but retains bounded-byte handling failure. The second defect computes record statistics using strlen instead of the public length: embedded NUL truncates ordinary content. It occurs outside the text-only default demo. only_fix_binary repairs length handling but retains the ownership defect. The defects are independent: each partial repair remains a compiled behavioral failure.

Gold uses a 37-byte prefetched source and geometric buffer growth. valid_exact reads one byte per callback and requests exact-fit allocation for each append; this changes both source data handling/control and allocation strategy, and deliberately has different read/resize counts. The checker does not demand those counts equal gold. noop preserves a compilable API while omitting normal processing; cheat masks digest to 16 bits; mask_symptom converts real terminal failures into success. cheat_read_grader executes real gold behavior only when opening a private check succeeds. Its trusted-host PASS is an allowed exception and does not establish host isolation; Docker must reject it.

Every behavior below is checked from raw public API/allocator observations against a Python model which never runs or calls gold. C harness code contains no hidden expectation or assertion; the P row reports direct API facts. All support, source copies and input are prepared before the first child. Thereafter parent check only uses memory and pipes, including finally cleanup via candidate-identity argv and an independently supervised strict os.lstat probe.

Criterion Exact public basis Coverage / observation
Offline clean compilation, normal program exit template/docs/contract.md:3 and :23; template/Makefile:2 compiler real rc and candidate real rc
API and result status agree; seed and zero initialization template/docs/contract.md:5 and :19 every J row; invalid maximum/stop/divisor/remainder
Ordered bytes, cache-independent delivery template/docs/contract.md:7 chunks 1/3/7/11/37/81, lengths crossing 37, delivered byte count
Empty, LF, consecutive LF, final suffix template/docs/contract.md:9 empty, one byte, newline pairs and suffix jobs
NUL, high bytes, CR are bounded content template/docs/contract.md:9 and :17 embedded NUL; bytes(range(256)); generated binary jobs
Exact maximum accepted and next byte rejects partial line template/docs/contract.md:11 1/2/15/16/17/37/38/127/256/1024 boundaries; prior line preserved
Error before/within/after complete lines; error wins over EOF template/docs/contract.md:13 offsets 0/1/3/4/5/7/8/9 with prefetch
Explicit allocator rejection yields LP_MEMORY; no partial publish template/docs/contract.md:15 and :21 reject_all on empty/text/binary jobs
No live allocations, live bytes or invalid release on any return template/docs/contract.md:15 and :25 EOF, stop, I/O, limit, invalid options, explicit allocation failure
Resize/release observations are nonnegative and conserve operations template/docs/contract.md:15 and :25; template/observe.c resize_block/release_block successes <= attempts; releases <= successes; successful nonzero allocation implies positive peak
Full line count/bytes/longest, selection by byte sum template/docs/contract.md:17 deterministic independent Python aggregate model
Rolling uint64 digest with per-line boundary template/docs/contract.md:17 seeds 31 and generated uint64, all byte values, complete lines only
stop_after prevents next-line delivery template/docs/contract.md:19 stop 1/2/3 with prefetch and generated jobs
Invalid options reject without callbacks/resources template/docs/contract.md:19 invalid max 0/65537, stop 1000001, divisor 0/257, remainder==divisor
Raw input and observation format template/docs/contract.md:21 and :23 one tag and exactly 18 ASCII decimal tokens per J row
Arbitrary signs/leading zeroes; arbitrary stderr permitted template/docs/contract.md:23 whole ASCII token validation then zero normalization; no stderr cap; production 17MiB/256-byte regression
Direct line append/content/sentinel, clear/reuse, destroy twice and invalid initialization template/docs/contract.md:15 and :27 P row exact defined 14 fields, no unused padding/cache checks

Necessary engineering counts are recorded per file in code-line-count.json: 683 ordinary nonblank noncomment lines, or 530 after additionally excluding standalone braces. No duplicate fixture, data, Markdown, Python oracle, or supervisor line contributes to that count. C sources alone have 599 ordinary lines; the public header and Makefile are necessary support. The observer includes byte source, ownership tracker, protocol decoding and API fact reporting.

Grader isolation uses byte-identical rollout/bos_verify.py and only check.py/context.json as grader_files. Candidate, compiler, cleanup and independent absence are child processes, with fixed production 20/10/5/5-second deadlines. The embedded supervisor inherits candidate UID/GID/groups, is a Linux subreaper, discovers descendants from current parentage or live same-start old tokens, polls the owned Popen before adopted waits, and freshly checks each signal PID/start/state. Z/X and reused old identities cannot become roots. ENOENT/ESRCH disappearance is normal; other observation/signal exceptions reject.

The parent validates native required fields, identity, exact integer types, finite bounded elapsed (deadline+5 seconds for reap), booleans, real leader rc/reaped, timeout/control rc consistency, absence, exceptions, IO lengths/SHA and command/nonce bindings. A wrapper nonzero, missing wire/field, preparation fault, blank OSError after any success flag, cleanup nonzero/timeout even after removal, present path or lstat exception cannot PASS. Raw wrapper length/SHA and actual wrapper rc are recorded before JSON parsing. Native records retain actual IO lengths/SHA; original inner wire and raw candidate IO bytes are not archived and are not claimed as preserved. Outer grade stdout/stderr are actual saved bytes with SHA. Native metadata is compressed to fit the bos 4000-character stderr tail and is decoded and checked from actual logs; a truncated record is explicitly incomplete.

Final diagnostics use actual complete core.load_task packages and fixed-image core._grade, not reduced deadlines. They include normal gold, 17MiB stderr cycling all byte values, leader 0 plus setsid/closed-stdio descendant reclamation, leader 0 holding pipes until production timeout, nonterminating leader, actual compiler FIFO, delete-then-nonzero/timeout cleanup, noop cleanup followed by present absence, absence timeout, strict lstat PermissionError, preparation/after-success/after-cleanup/after-absence empty OSError, missing all wire metadata, missing leader_start, supervisor nonzero, boolean elapsed, negative IO length, and long plus/leading-zero stdout. Earlier-generation diagnostics are retained separately and are not described as final-generation coverage. Pure synthetic /proc tests do not involve real PID exhaustion or unrelated processes.

The local and Docker gates each execute all nine fixture classes five times with unique attempt directories. Actual paths, task digest, image identity, per-class verdicts, native decoding and cleanup checks are in report.json. No Qwen/Pi/real Agent or hardware run is authorized in this task scope. Root runs the repository-wide suite; the author's nine focused behavior tests have original logs. Final independent review and any later difficulty sampling remain outside this delivery.