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

14 KiB
Raw Permalink Blame History

numeric-unit-profile / version 1

Status CALIBRATED; difficulty UNMEASURED. Route C original software fixture, source hwe/dbg-gdb-frame-const, split key dbg-gdb-frame-const. Author is the assigned independent task author; root review is pending. HWE prompt and solution were read only; no source scripts executed. Original text misstates integer 9/5 as zero: it is one, so its original claimed symptom is not reused. This independently authored profile service retains the premature integer-ratio calculation knowledge and tests a second independent reverse affine-order defect.

Software substitution preserves deterministic numeric behavior without ARM, debugger interaction, CMake, real timing, network or hardware. It does not claim a historical field incident. Business code, public observer and independent Python Fraction model are newly authored, with only the root-reviewed execution-only attachment reused for process supervision. No other batch business task was read.

Defects: D1 forward gain truncates numerator/denominator before multiplication, triggered by the default forward demo (9/5). D2 reverse uses y*d/n-offset rather than (y-offset)*d/n, triggered solely by the optional reverse example, absent from default demo. only_fix_ratio fixes D1 alone; only_fix_inverse fixes D2 alone; both must fail. gold fixes both. valid_rational uses a different exact-rational arithmetic organization. Additional noop, hardcoded cheat, reader-gated gold, mask_symptom are behaviorally rejected.

Candidate files src/unit_profile.c and src/unit_registry.c; protected grader only check.py; verify.py byte-copy of rollout/bos_verify.py. Public interface/header, observer and input examples are frozen as check strings before any candidate child. Candidate compiled with raw observer; expectations remain parent-only. Fixed-seed cases cover all public clauses. Full per-clause file/line matrix and actual evidence will be added after source publication, before freezing.

Production deadlines planned compiler=20s, candidate=10s, cleanup=5s, independent absence=5s, recovery=3s; total task limit=90s with verify inner limit=100s. Candidate-identity subreaper closes same-PG and setsid descendants, polls/waits own Popen before adopted wait, uses fresh PID/start nonzombie identity before signals and bounded retry on exceptions. Unknown adopted wait start stays None. Parent validates exact types/bindings/IO digests and rejects any exception, timeout, nonzero or missing metadata. Strict independent lstat follows cleanup and is separately required. stderr is arbitrary bytes without size cap; public integer output token parser validates full ASCII tokens and normalizes arbitrary leading zeroes.

No Agent/Pi/Qwen sampling, no difficulty measurement, no hardware or full-library suite is authorized for this single-author task. Only local and fixed-image per-class five-repeat calibration after diagnostic proof and final freeze. Successful helpers/proofs/gates will not be rerun to repair collectors. All actual source/check/wire/IO evidence and exclusions are reported honestly; formal tail metadata compressed for full decoding, IO not retained is only actual length/SHA.

Exact public basis and independent observation matrix

All paths below are relative to template. Every numeric expectation is independently computed with Python Fraction from these public rules, never by executing gold. The raw observer is compiled unchanged with candidate C files. Only exposed fields/valid element ranges are compared; padding and unused tails are ignored.

Criterion Exact public basis Covered observation
C11 offline public make and CLI docs/unit-profile.md:3; Makefile:1 Fixed-image gcc14.2/make4.4.1 probe; five public package builds and both examples exit0
Profile id validity docs/unit-profile.md:4 D/X/Z/P/S/C/B/Y invalid 0/256 and valid 1/255
Nonzero signed gain and numerator bounds docs/unit-profile.md:4 D zero and ±1001 reject, valid negative/positive gains
Denominator bounds docs/unit-profile.md:4 D 0/1001 reject; valid 1/2/3/7/13/999/1000
Offset bounds docs/unit-profile.md:4 D ±1000001 reject; ±1000000 accepted
Base endpoint bounds/order docs/unit-profile.md:4 D out-of-bounds/inverted endpoints reject; singleton domain accepted
Precision domain docs/unit-profile.md:4 D -1/4 reject; 0/1/2/3 exercised
Invalid definition is atomic docs/unit-profile.md:4 P/S/L after every invalid replacement preserve configuration/statistics/revision
Capacity/full and replacement docs/unit-profile.md:4 Six occupied, seventh FULL; overwrite at capacity resets statistics
Definition revision docs/unit-profile.md:4 L after successful/failed D counts exactly successful changes
Exact forward scale plus offset docs/unit-profile.md:5 Default 9/5 Celsius; multiple signed rational gains and offsets
Exact inverse offset before reciprocal gain docs/unit-profile.md:5 Independent optional reverse; generated signed/offset combinations
Quantization after full affine map docs/unit-profile.md:5 All four decimal settings; no intermediate rounding; exact Fraction oracle
Half ties away from zero docs/unit-profile.md:5 ±1 at gain1/2, additional signed half-bin cases
Argument validity and precedence docs/unit-profile.md:6 Invalid direction/id, missing id and empty batch combinations; H raw input-limit helper
Missing profiles docs/unit-profile.md:6 Initial missing C/B/P/S/X/Z/Y and post-delete queries
Inclusive forward domain docs/unit-profile.md:6 Endpoint, endpoint±1, singleton domain
Unrounded inverse domain docs/unit-profile.md:6 Exact recovered boundary and just outside before rounding
Rounded output range docs/unit-profile.md:6 Gain1000, offset1000000 at base1000000 yields RANGE
Failed scalar output is untouched docs/unit-profile.md:6; docs/unit-profile.md:12 C caller sentinel and H output printed even on failure
Successful sample statistics docs/unit-profile.md:7 S accepted/min/max/sum/last and ordered batch outputs
Failed sample rejection accounting docs/unit-profile.md:7 One rejection for existing id, no successful output changes
Empty statistics and query stability docs/unit-profile.md:7; docs/unit-profile.md:9 S before conversion, after clear/replace/reset, interleaved queries
Batch atomic first failure docs/unit-profile.md:8 Failing middle element leaves every output at sentinel; one rejection, no partial acceptance
Batch success and maximum count docs/unit-profile.md:8 0/3/4/16 elements; all output elements/order and statistics
Batch pointer/count validation docs/unit-profile.md:8; docs/unit-profile.md:10; observe.c:12 H null input/output and count17; empty null/null succeeds
Erase/reset/clear revision and occupancy docs/unit-profile.md:9 Repeated empty reset, delete, missing delete, independent stat clearing
Sorted enumeration, bounded copies docs/unit-profile.md:9; docs/unit-profile.md:10 Random six-id L plus H capacity0/1/null probes
Summary uses occupied entries only docs/unit-profile.md:9 L before/after deletion/duplication/reset, accepted/rejected totals
Duplicate atomicity and fresh statistics docs/unit-profile.md:9 Missing/invalid/full, occupied replacement, self-duplicate and new slot after delete
Exact configuration lookup docs/unit-profile.md:9 P after every define, invalid replacement, duplicate and erase
Null outputs/definitions and pure helpers docs/unit-profile.md:10; include/unit_profile.h:50; observe.c:12 H prints each actual return, sentinel, statistic and partial enumeration
Well-formed command input domain docs/unit-profile.md:11 Frozen input commands use public decimal fields and space/LF delimiters
stdout integer lexer and field protocol docs/unit-profile.md:12 Whole ASCII token validation, sign/leading-zero normalization, exact rows/fields; 5000-digit zero helpers and actual 5000-width numeric output proof
Arbitrary stderr accepted docs/unit-profile.md:11 Actual all0..255 and 17MiB all256 diagnostics PASS without stderr cap
Raw H observation order docs/unit-profile.md:13; observe.c:12 Parent-only independent expected values for public edge probe

Actual author verification before freeze

Ordinary unique engineering counts exclude blank/comment lines, fixtures, observer, docs, examples and private grader: template 503 = profile138 + registry281 + header71 + necessary Make13; gold501 =136+281+71+13. This is the single engineering count required by this batch; no additional candidate-only500 claim. Library responsibilities include exact affine evaluation/validation, six-entry lifecycle, duplication, atomic batches and output statistics. Header/Make are necessary to build the public engineering project. valid_rational is intentionally shorter, sharing the registry but replacing arithmetic with reduced rational composition and binary search for a quantization bin, rather than gold's quotient/remainder rounding.

Final deterministic matrix contains 2384 command/observation rows and 11074 exposed integer observations; payload34492B, canonical expected stdout42059B. H is a public fixed edge-observer; other cases include fixed Cartesian signed gains, four precisions, domains/rounding/range, six-entry transitions and seed2026102503 random registry sessions. After the first 22 process diagnostics, one generator-only change added valid16-element batches (32 additional commands). Imports, constants, supervisor, parent stage execution/validator, public observer, cleanup and absence code remain byte-identical; the three subsequent diagnostics use the final matrix. Older diagnostic packages and exact digests are retained and are not represented as current-package re-executions.

25 immutable full-package fixed-image core._grade diagnostics all met expected verdicts: normal gold; stderr256/17MiB all256; normal leader0+setsid closedstdio; compiler FIFO20s; candidate fork10s; leader0 held-pipe10s; cleanup delete+exit7/delete+timeout5s; cleanup noop with independent absence1; independent absence5s timeout; strict lstat empty OSError and real EACCES; preparation empty OSError; errors after business success, cleanup0, absence0; actual wire missing metadata/field, UID bool and nonnull adopted start; supervisor exit9; fresh signal empty OSError with actual later retry; candidate exit7; legal long-leading-zero stdout. All have full actual package/check/context/candidate hashes and raw wrapper wire/stdout/stderr in a protected diagnostic mount. Normal closedstdio records leader0 and adopted [PID,None,-9]; error retry records leader0, retained empty OSError, one actual signal and adopted [PID,None,-9], closure/PG empty, still FAIL. No historical tracked starttime substitutes for adopted wait identity.

Author helper tests: 8 original arithmetic/token/closure tests PASS, then 5 saved-metadata/type/fresh-signal tests PASS (the latter binds before/after source and logSHA). Original8 helper log/source are retained but had no contemporaneous before/after binding; do not infer one. No successful helper/proof/main was rerun. Fixed-image public make/demo/reverse proof covers template,gold,only_fix_ratio,only_fix_inverse,valid_rational: all build/run0; gold default equals only_fix_ratio and differs from only_fix_inverse; reverse has the opposite equality. D2 is absent from actual default demo, not just labeled nondefault.

Process proof has explicit limitations: metadata describes supervisor identity and observed inherited leader lifecycle, not a separate post-exec credential read. Input payload is captured by the controller but there is no candidate consume acknowledgement. Diagnostics retain actual full wire/IO; normal calibration retains complete compressed stage metadata and actual IO length/SHA only, not original wire or IO bytes. Do not reconstruct missing originals and label them retained. Synthetic proc graphs cover stale/reused/Z/X roots and fresh-signal identity; real PID exhaustion is not performed. No hardware/debugger navigation/ARM performance, network, real Agent, Pi/Qwen, sampling, full-suite tests, git changes or platform edits are claimed.

Calibration receipts and final digest will be recorded in this spec/report after the two authorized gates; independent root review is pending.

Final gate facts

Final task digest 0ddddb1c5c77a7e6194df2ae24193ab0e246a37d568d7f840b33a2f4339c00c9; checkSHA d65493e22c1cd859fa898a987018b8c0a05f9acb9f3d3039510fc5f956d79bb5. One local45 and one Docker45 completed, both overallPASS and cleanup confirmed, all nine classes five times each as expected (readerFAIL in both; no host isolation claim). Actual receipts: /home/ming01.wang/embedded-rollout/.tmp/wt-batch15/runs/20261023-C-hwe-scale-15/dbg-gdb-frame-const/local/calibration-f39e4947c0554d4bb8573571a75c8097/receipt.json and /home/ming01.wang/embedded-rollout/.tmp/wt-batch15/runs/20261023-C-hwe-scale-15/dbg-gdb-frame-const/docker/calibration-4dd84abb4b514e508cb73736e8f58abd/receipt.json. Full360stage records decoded from90 logs;350leader0 and10reader71;20positive actual stdout42059B length/SHA matches explicitly labeled expected reconstruction. Original formal wire/IO not saved. Exact72owned containers,363local observedPID/start identities and45local scratch paths freshly absent. Full source/spec/report/rootreview pending; stop for independent review. One pre-freeze trusted helper import failure retained in report with original sourceSHA and tool-observed error text, without claiming a retained raw stderr file. 族名规范化:numeric-unit-conversion → numeric-reliability(2026-10-27)