# 0.6.15 candidate: four native platforms

Release PR 1396: https://github.com/kimjooyoon/meta-ontology-go/pull/1396

Candidate `a202e6e5b747cb14ae60ee82252929aec6755c0a` passed all four native
platforms, aggregate replay and 7/7 eligibility in readiness run
https://github.com/kimjooyoon/meta-ontology-go/actions/runs/37835432213

Publication conformance run 37835432303 also passed. Canonical PR CI 37835432048
failed semantic conformance: automatic extraction of the new validation function
reported `NO_SAFE_DECLARATION_CAPACITY`. The original failure log is retained.
The same checks were separated into smaller functions using the existing named
input type; original cases and compiler construction semantics are unchanged.
Local release tests, 31 changed-evidence cases, race and vet passed again.

The repair head is `9c5c1e37bf22da1ab8f6046d5b35c82085563438`. Readiness 37836963879
and publication conformance 37836963706 passed. Canonical CI 37836963810 is still
running; its semantic-conformance job now passes. Final eligibility artifact
11576346430 (1,048 bytes) matched
`sha256:4ef5bc7d06bccabc8602bbe57b90f760b587707c8c7a60cadc39399d29b77c52`.
The exact source and 7/7 closed eligibility were checked. The compact raw
four-platform comparison archive in this directory describes the earlier
candidate above; the new aggregate was not downloaded. This is pre-release
evidence; 0.6.15 has not been published or installed.

The original aggregate ZIP artifact 11575961395 (160,949,145 bytes) matched
`sha256:00df353d13928160b38174eb6c0c35d47beb8532a980fdbc2b47a3d4c5aad013`.
Only the four native receipts, twelve new rejected-fill raw outputs and eligibility
were retained in the smaller archive here. The original ZIP was removed after
verification. `FILES.sha256` describes the retained subset, not the original ZIP.

`compare-platforms.go` takes the extracted subset directory, the local candidate
observation directory and the expected CI SHA. It decoded exact JSON numbers and
checked all four clean native build identities, platform OS/architecture, archive
replay, five original assignments, rejection records, source/plan IDs, budgets,
original case values, selected source, partial 1/4 and complete/replay 4/4 results.
All twelve records matched the clean macOS candidate `848f85de` observations.
The later source changes only store documentation/evidence. These are repeated
platform executions of the same examples, not twelve independent programs.

The compiler repository retains the frozen plan, actual fixtures, local evidence
and original cases under `docs/research/release-0615-20261009` and
`examples/caller-fill-rejection`. The workbench recount helper `observe.go` here
is run from the workbench module, passing a directory of the six gzip outputs and
original evaluation files. It preserves exact numbers and checks selected-source
equality across fixed/model ordering and replay. No new model was trained.