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

9.5 KiB
Raw Permalink Blame History

state-step-progress

Route C ORIGINAL. Source hwe/dbg-gdb-infinite-loop-inc; base/split group dbg-gdb-infinite-loop-inc; family state-machine; C11; version 1. Source material is an easy/public missing-increment demonstration, not a proven field incident. Its CMake requirement is background only; this independent software project uses the fixed runtime's verified gcc/make. No source scripts were executed, no other batch business code was read.

The software substitute retains the original reliability issue: a loop must update the work-domain control state so that a finite sequence of work completes. A finite transition budget reports BUDGET instead of allowing a real infinite loop. The project is a useful pausable filtered scanner with a raw transition journal, reset/replacement, counters and an order-dependent result digest. It does not test ARM code generation, hardware, gdb command use, or real timing. Difficulty is UNMEASURED; no model trial was run. Status is CALIBRATED pending root review.

The initial spec was recorded in spec-initial.md before implementation. Public supporting engineering is newly written: validated command parsing, dynamic LF input, journal accumulation, signed selection/digest, raw formatting and Make examples. Only scanner.c is candidate material; support and the observer are frozen into the check string. The check uses an independent Python closed-prefix model: an immutable domain index sequence, accepted prefixes, and arithmetic mapping from total transition ordinal to state/event. It never runs gold to calculate expected values. All source/support and all stdin bytes are prepared before the first child; after that parent code uses only memory/pipes and protected diagnostic evidence, with candidate-identity scratch cleanup plus independent strict-lstat absence in finally.

Independent defects

A: committing an excluded record retains cursor. Default make demo reaches an excluded record and repeats it under the controlled budget. Public clause 3 requires every commit to advance, independently of selection. only_fix_reverse retains A; only_fix_progress removes A.

B: a one-sided terminal comparison treats the starting index of a reverse range as complete. This is independent of A: the reverse example selects every record, so A is inactive. The actual default build/demo loads only the forward demo. The optional make reverse example is separate, documented and independently invoked. Fixed-image gold and only_fix_progress produce matching default behavior; only_fix_progress fails the reverse example. Public clauses 3 and 8 require reverse traversal. only_fix_progress retains B; only_fix_reverse removes B.

Public evidence mapping

Checked behavior Public exact location Coverage
Initialization and unconfigured errors template/docs/scanner-contract.md:3 S/R/T before P
Valid and invalid plan fields; replacement and empty/singleton plans template/docs/scanner-contract.md:5 P plans including boundary values, malformed P preserving prior state
Forward and reverse inclusive domains; unused array values not scanned template/docs/scanner-contract.md:7 random independently selected first/last and all lengths
Three transitions and exact paused state template/docs/scanner-contract.md:7 budgets 0,1,2 and varied continuation
Excluded records still advance and commit visit count template/docs/scanner-contract.md:7 mixed/default and all-excluded filters
Signed modulo and accepted totals template/docs/scanner-contract.md:7 negative and extreme values, modulus 1/17
Position/value-dependent digest template/docs/scanner-contract.md:9 independent accepted-prefix polynomial
Finite budget, completion boundary and repeated DONE template/docs/scanner-contract.md:11 oversized/exact budgets and repeated T
Raw events active prefix and before-advance index template/docs/scanner-contract.md:11 every generated transition event compared
Strict input tokens; allowed signs/zero padding; exact separators template/docs/scanner-contract.md:13 empty/malformed/extra/VT and 5000 leading zeros
Raw stdout schema; horizontal whitespace and numeric equivalence template/docs/scanner-contract.md:15 decimal parser plus legal longzero-output diagnostic
Arbitrary stderr does not alter accepted behavior template/docs/scanner-contract.md:15 all 256 byte values and 17MiB actual stderr
Restart and optional reverse example outside actual default demo template/docs/scanner-contract.md:17 fixed-image public make+demo+reverse of gold and only_fix_progress

Fixtures

  • gold: Switch-based transition engine, PASS.
  • template: Excluded commit does not advance cursor; reverse terminal comparison stops at first item. FAIL.
  • noop: Tick emits no transitions and leaves work unfinished. Builds, FAIL on observations.
  • cheat: Accepts every record by replacing the supplied filter. Builds, FAIL on totals and trace.
  • cheat_read_grader: Gold behavior only after opening /grader/check.py or /grader/verify.py. Otherwise tick returns -1. Docker denies protected reads; trusted local result is not an isolation proof.
  • only_fix_progress: Cursor progresses for excluded records; reverse completion comparison remains wrong. FAIL only on reverse scenarios.
  • only_fix_reverse: Inclusive reverse completion fixed; excluded records still repeat. FAIL on forward mixed/excluded cases.
  • mask_symptom: Changes stop to DONE whenever a positive budget arrives, without doing the transitions. Builds, FAIL on raw counters/events.
  • valid_table: Alternate accepted-prefix arrays plus absolute transition ordinal reconstruction. Recomputes prefix counts/sums/digests and derives each observation by division/modulo, rather than the gold switch/staged mutation. PASS; historical fixture name remains valid_table.

Execution verification and limits

The authorized same-task execution attachment supplied only execution machinery. It was adapted to selector-based uncapped byte IO, honest adopted [PID,None,exitcode] records, strict identity/boolean/finite elapsed/IO lengths and digests, fresh same-start non-zombie signals, bounded finally closure recovery, own Popen poll/wait before adopted reap, strict lstat and independent cleanup/absence. Each stage is executed as UID/GID65534/groups[] in Docker. Compiler/candidate/cleanup/absence deadlines are 20/8/5/5 seconds; recovery is 3 seconds. Cleanup nonzero, timeout, missing metadata and any exception reject even when work or absence succeeded.

26 complete actual-package trusted diagnostic copies were run through fixed-image real core._grade. They include normal gold, all256 and 17MiB stderr, normal leader0+setsid/closedstdio child PASS; real compiler FIFO20s and leader0 held-pipe8s timeout; cleanup delete+7/delete+5s timeout with independent absence0; no-op cleanup0+absence1; strict kernel EACCES and absence5s timeout; preparation and post-success/cleanup/absence empty OSError; a live nonleader freshstat empty OSError followed by retry/signaled/adopted cleanup yet FAIL; incomplete wire, supervisor7, inconsistent types/nonfinite elapsed, nonnull adopted identity, and legal 5000-leading-zero raw output. Every diagnostic has the actual loadable package, before bindings, original wire bytes, raw logs, decoded metadata and actual IO digest verification.

Only 8 local pure tests were initially run: seven passed and one hand-written expected digest was mistyped. The original source/log/error binding remain intact; only that failed assertion was corrected and independently rechecked. No successful helper was rerun. Synthetic proc tests cover stale/reused/dead roots, live remembered roots, fresh-signal authorization and unknown adopted start None. They do not exhaust real PID reuse or kernel scheduling states.

The formal bos tail stores one complete compressed metadata record, actually decoded after every run; original large formal wire/IO are not saved. Formal records have actual byte lengths and SHA256, and any canonical oracle reconstruction will be explicitly labeled as reconstruction. Diagnostic wire stores the actual losslessly compressed IO. The recorded leader identity is captured just after Popen, not an independent post-exec credential handshake. Stdin supplied to a pipe does not prove candidate consumption. A missed/reaped adopted start is deliberately None and does not authorize signaling. Local read-grader behavior does not demonstrate host isolation. No global unit suite, platform edits, git operations, real model sampling, hardware tests or ADMITTED claim are included in this subtask.

The original digest d041cd75dce031a74c36be7ec985c16b973da25aa6535cafe84ea237491a28d4 had 90 CLI-PASS trials but its noop sample failed compilation in ten trials. Those gates are retained and excluded from final fixture-quality admission. Only noop/scanner.c changed; check, gold and public context stayed byte-identical. The new digest receives its own complete 90 trials; successful unaffected diagnostics and helpers were not rerun.

Final calibration receipts

  • local: /home/ming01.wang/embedded-rollout/.tmp/wt-batch15/runs/20261023-C-hwe-scale-15/dbg-gdb-infinite-loop-inc/r1/local/calibration-de8766114f6443309c2e8a4e632ab04b/receipt.json, PASS, 45/45 expected outcomes, cleanup confirmed.
  • docker: /home/ming01.wang/embedded-rollout/.tmp/wt-batch15/runs/20261023-C-hwe-scale-15/dbg-gdb-infinite-loop-inc/r1/docker/calibration-d071e2781f48493ea6730dafff38c2b3/receipt.json, PASS, 45/45 expected outcomes, cleanup confirmed.

族名规范化:state-machine → state-machine-timing(2026-10-27)