ci: route exact limited timestamp release branch to Stage
Deploy Canina / deploy (push) Successful in 2m51s

This commit is contained in:
Canina Hermes committed 2026-10-30 13:28:55 +00:00
1 parent c97ea01ba9
commit dc34c57a0c
4 files changed
+131 -3

No files matched your search

@@ -0,0 +1,22 @@
# Limited release Stage routing — 2026-10-30
Scope: preserve mixed `develop` and its historical evidence. Add only the explicit `release/blog-timestamp-limited-20261030` push ref to the official deployment workflow. No push or deployment was performed by the preparing worker.
## Source contract
- Workflow binds `DEPLOY_REF` to full event `github.ref` and leaves `DEPLOY_SHA` bound to `github.sha`. Unknown refs and malformed SHAs fail before downloads.
- Only `refs/heads/main` selects `deploy.sh main`; develop and the exact dedicated release ref select `deploy.sh develop` with an explicit Stage source ref.
- `deploy.sh develop` execs `release-cache-stage.sh`. That script previously fetched only develop and required `origin/develop == DEPLOY_SHA`; workflow routing alone was therefore insufficient.
- Stage now fetches the exact allowlisted dedicated source into its own remote-tracking ref and requires its SHA to equal `DEPLOY_SHA` before archive/build. The develop path remains unchanged. No force refspec or operational checkout reset is introduced.
- `deploy.sh` and `release-cache-prod.sh` are unchanged. Production still fetches main and requires `origin/main == DEPLOY_SHA`. Both official release scripts are code-only, building/recreating only backend/frontend through the existing lock/runtime/provenance checks; no migrations added.
## Actual local execution
- New source-executing regression suite was run RED before implementation: missing allowlisted push ref, missing Stage source-ref guard, and missing fail-closed full-ref routing failed.
- `python -m unittest scripts.tests.test_limited_stage_route scripts.tests.test_cache_stage_release scripts.tests.test_cache_release_safety -v`: **13 tests passed**, 17.845s. Includes actual embedded workflow shell preamble execution, exact/near-match/tag/injection rejection, invalid SHA rejection, Stage source-ref guard with a fake Git transport, mismatched remote SHA rejection, and actual deploy dispatcher execution with isolated release doubles.
- `bash -n scripts/deploy.sh scripts/release-cache-stage.sh scripts/release-cache-prod.sh`: passed.
- `git diff --check`: passed.
- Additional `test_production_code_only_ci` run: 2 passed, 1 failed because the baseline production script already has `--repair-empty` while an old test forbids it. Reproduced that exact assertion using `git show 00ba4360:scripts/release-cache-prod.sh`; production source is byte-identical to baseline. No production script/test was changed to hide that existing failure.
- Local Graphify inventory was queried: deploy.sh nodes exist but their old locations describe an obsolete script and no current release-cache-stage/workflow nodes were present. Actual source above is authoritative; no claim of a fresh graph or graph-derived release acceptance.
Existing timestamp test files and evidence were preserved byte-for-byte. This evidence is local routing verification only, not official Stage acceptance or Production acceptance. Parent remains sole push owner and must verify CI/run SHA, immutable images, and live timestamp behavior before promoting the identical accepted SHA to main without a merge commit.