Stabilize packaged release build graph

This commit is contained in:
Meta Vens 2026-08-09 05:46:51 +01:00
commit fe6bf006d3
5 changed files with 138 additions and 47 deletions

View file

@ -1923,3 +1923,25 @@ vendored Starlark's older `hashbrown` integration; Rust rejected the missing
release-target check traversing Starlark through Codex login and
`epiphany-openai-auth-spine` completed successfully in 3m16s. This is lockfile
compatibility authority, not a source patch or relaxed gate.
### Single deterministic packaged build authority
The original packaged-release path invoked Cargo three times against one target
directory: the broad root binary graph, the coordinator feature graph, and the
core-owned state steward. Cargo correctly rebuilt when each command changed the
selected feature set, so the cache oscillated instead of warming. Exact source
`e9465c11` took 24m22s cold (17m03s root plus 6m38s state) and 28m30s on an
identical cached rerun (21m04s root, 34.62s coordinator, 6m42s state).
The two runs also yielded different valid authenticated releases. Twenty-one of
23 binaries were byte-identical; only the two `epiphany-openai-runtime`
binaries differed, including different ELF build IDs and code bytes. Release
incremental codegen is therefore not an admissible input to witnessed output.
`build_required_release_siblings` now has one process owner. It selects
`epiphany-release-bundle` and `epiphany-core`, selects
`epiphany-release-bundle/coordinator-runtime` once, names all 23 required
binaries explicitly, and sets `CARGO_INCREMENTAL=0`. Follow-up coordinator and
state-steward Cargo commands are forbidden writers and have been deleted. All
19 focused packaged-release tests pass. Exact Linux packaging plus a
byte-identical warm rerun remain the verification boundary.