# Native process phases and failed output, 2026-10-09

Compiler source `aa79dc9462365d2049da5294e8a0917ad2257fee`,
[PR 1411](https://github.com/kimjooyoon/meta-ontology-go/pull/1411), is independently
based on dev `13d576ef`. Package-interface PR 1410 remains separate and open.
The candidate was built from a clean checkout with Go 1.27.2 and decision runtime
0.2.26. The installed public binary remains unchanged.

## What changed

Process evidence adds optional Start/Wait intervals, the original remaining
deadline, a context snapshot when Start returns, and the host operation reporting
the failure. Their sum equals the original total. Successful Start still calls
Wait, with the same contexts, two-second native limit, one-second pipe wait delay
and cleanup. Legacy observations retain absent phase values rather than gaining
invented zeros. Cache copies own their timing values.

The language release helper now saves original failed stdout/stderr bytes before
returning the command error. `.failed-output` distinguishes incomplete/diagnostic
bytes from successful JSON, including empty output. A persistence error retains
the original command error too.

Old Windows timeouts at candidate 8951c5f8 remain unexplained by their old data.
These changes supply additional measurements; they are not proof of a root-cause
fix. No old run was restarted and no timeout extended. PR CI and four-platform
readiness were running when this observation was published.

## Checks

The two red logs show missing process-phase fields and missing failed-output
capture before implementation. Focused green logs cover failure causes, invalid
diagnostic bytes, output bounds and cancellation. The full race run covers both
affected packages: bodyexecution 162.979 seconds, toolchainrelease 12.492 seconds.
Tests cover phase sums, expired/canceled starts, exit status, old records without
timing, independent cache copies, successful JSON paths, empty/invalid failed
output and simultaneous execution/persistence errors.

The actual packaged-compiler native caller profile passed in 14.952 seconds.
It includes eight construction/input/replay/evaluation records and exact plain
outputs `3`, `-10`, `-9007199254740994`, `1`. Affected-package vet, whole compiler
build, formatting and diff checks passed.

## Own-model use

`dogfood-original.tar.gz` stores byte-original fixed/model execution and saved
replay, manifests, actual inputs, Gooo policy source, timings, hashes and value
comparisons. `dogfood.sh` records the calls and assertions. The policy and three
input changes are the same as the workbench's published API-change experiment.
Actual classifications exactly match the prior compiler 8951c5f8 observation.

The existing model SHA-256 is
`3c68205a660695103712e2115ce90accd298b406dced8ad5ad5c9b84bab0a202`.
Neither weights nor source were modified. Fixed/model generation calls were
0/1, both attempted eight combinations, and both checked 51/51 selection fields.
The three actual input rows had no expected labels, so actual score stays 0/0.
Saved replay introduced zero new inference. All observed phase sums matched.

| Whole package execute | Elapsed s | User s | System s | Peak RSS bytes |
| --- | ---: | ---: | ---: | ---: |
| Fixed order | 0.63 | 0.20 | 0.11 | 86,917,120 |
| Own model | 0.40 | 0.22 | 0.11 | 86,687,744 |

Model prediction was 220,834 ns; record generation was 19,754,541 ns. The model
run's first child had Start 867,666 ns and Wait 193,571,250 ns; its second child
had Start 2,511,125 ns and Wait 5,492,667 ns. Both Start-return contexts were
ACTIVE. Wait includes runtime startup, scheduling and I/O as well as execution.
These intervals do not identify when Gooo user code started.

This is one local sequential sample per mode with differing cache/order conditions.
There is no controlled speedup, host CPU utilization or general correctness claim.
No new model was downloaded or trained.

## Workbench post-merge check

The retained run/job metadata confirms workbench main run 37882197317 succeeded
at `fd53420e3851c46870f846086bc152418f936d4b`. The previous observation contains
the original final PR artifact verification and candidate/merge tree comparison.
This metadata closes the previously pending post-merge status; a second copy of
the complete post-merge artifact was not downloaded.

## Subsequent Windows and readiness evidence

PR readiness run 37883255799 completed successfully at aa79dc94 on all four
platforms, including Windows job 113667493608. The aggregate passed and the
original final artifact 11595550600 (1,046 bytes), SHA-256
`5baf1a7068f4e511c9e31c13ed6a3565018271ca4e8fde125a1f98dab4d5428f`, was checked
against API metadata. Its exact-head receipt reports EVIDENCE_CLOSED/EXACT, 7/7
closed, no unknown/refuted work and zero repository writes. The original ZIP,
receipt, metadata and verification are retained under `readiness-final/`.
This is PR readiness evidence; canonical CI 37883255852 was still running.

Original Windows artifact 11595094224 (14,151,979 bytes) had SHA-256
`2bec80fdd908d7665a3e16a9df6cda33a0e6f180f96fde0fb6a0ee3c4fc5f249` and was
downloaded and digest-verified. `windows-original-json.tar.gz` preserves every
original JSON file extracted from it. The full ZIP, which also contains the
binary archive, remains in GitHub Actions and the local work record. Metadata
and extracted timing summaries are retained alongside the JSON archive.

The JSON corpus contains 481 appearances of process records, including build
records and repeated history/cache references. The native `runs` arrays contain
198 appearances; these are not counts of independent executions. Every observed
Start+Wait sum equals its total, and every Start-return context was ACTIVE.
The largest native Start interval was 171,646,600 ns in the caller-fill-rejection
partial record; its Wait was 6,004,000 ns. The largest native Wait was 17,842,500 ns
in the caller-fill partial record, with Start 71,925,300 ns.

Text-operations saved replay now completed 12/12. Its two native process totals
were 67,820,200 and 7,650,200 ns. The first Start took 61,471,600 ns and Wait took
6,348,600 ns. This successful observation provides phase measurements. The older
failed observations lacked those phases, so their cause remains unconfirmed.

The original failed text-operations replay and this successful replay have equal,
nonempty executable, generated-code, driver and input-suite digests. Both executed
binary SHA-256 `21e12aef61580a336f0e86cf9aa238277244eaa51e4f1bea802676e9d643f2ea`.
`failed-vs-completed-identity.json` and `identity-verification.txt` retain the
comparison against the original failed JSON published in the previous observation.
This narrows the investigation to execution conditions and deadline handling for
the same program/input. It does not identify a specific host component as the cause.

## Later completion and separate ledger failure

Canonical run 37883255852 subsequently completed successfully. Its original
artifact 11596430789 (11,164 bytes), SHA-256
`208fd48a2c604426805c4a1f1a9f850e9abcf38dc5dbe2fac7e1ad26ae05eae7`, passed
the six-job, source/ref, digest and provenance checks. PR 1411 was normally
merged with the expected head to dev `31d3419cd1b8a66d01fa1a2115c50831b0898944`;
the merged tree equals the tested aa79dc94 tree.

The full PR check listing also revealed a separate Windows ledger failure in
run 37883255912. Its process Start consumed 2,854,721,800 ns of a two-second
deadline; Wait then took 510,000 ns. The original phase-observation change was
integrated into dev with this known failure retained. This was not a main
promotion or release, and the PR was not described as fully green.
The original proof, merge metadata, failed artifact and subsequent budget
correction are preserved in [the follow-up observation](../native-wait-budget-20261009/README.md).