Keep auxiliary screenshot work from displacing shared native caches

Status

Operational change applied and read back on 2026-09-28. This note records the upstream workflow registration and cache cleanup; committing documentation alone does not change GitHub settings. No branch-protection rule was changed.

Context

The required native suite publishes reusable default-branch snapshots and restores them read-only on PRs. An auxiliary screenshot workflow introduced through open fork PRs independently built complete Windows/macOS release applications and published large PR-scoped Rust caches in the upstream repository.

The screenshot workflow is absent from main at 8087306b, but its GitHub registration still accepted events for PR branches that supplied the file. Editing only a workflow on main would not change those already-open fork branches.

Evidence

Native run 36327240193 took 85m29s. Its Intel Mac queued for 48m07s and executed for 36m47s. Both Mac architectures missed their native and release target caches, with the same keys used by previously available default-branch entries.
The cache inventory contained four v0-rust-native-ui- snapshots scoped to PR merge refs, together using 5.39 GB. The previously observed default-branch Mac target snapshots were no longer present. These observations establish cache pressure and misses, not an audit of who deleted each older entry.
Screenshot run 36325527329 used .github/workflows/ui-review-screenshots.yml from a fork PR. Its source ran on both fork pushes and upstream PR updates; its publication job was explicitly restricted to the source fork's push events. Upstream repeated the full application builds but did not run that publication job.
The screenshot action used Swatinem/rust-cache with failure caching and no read-only restriction. GitHub's cache scope rules do not make one PR's merge-ref cache available to sibling PRs merely because their visible cache keys match.

Decision

Disable only the upstream UI review screenshots registration, workflow ID 366570225, through the documented workflow API. Leave source-fork workflow 366585525 active. No contributor branch or application code is rewritten.
Preserve the full five-platform native suite, release-profile checks, required contexts, assertions, feature graphs, packaging and branch-protection rules. Auxiliary screenshot evidence remains separate from native regression coverage.
After confirming there are no active screenshot runs, remove only identified upstream PR-scoped v0-rust-native-ui- entries. Preserve every default-branch cache, release artifact and source file. Do not install recurring broad cleanup.
Keep contributor screenshot publication in its source repository. Any future upstream screenshot entrypoint must be explicitly requested, restore target caches read-only, and retain exact-source evidence before it is enabled.

Rejected alternatives

Delete all caches: would force more cold native and release checks.
Raise storage quota or buy runners first: does not address duplicate upstream builds or the unshared snapshots that caused the observed competition.
Push fixes into all contributor branches immediately: creates new PR heads and repeats their required native matrices. The upstream registration can be stopped without editing their feature code or interrupting their fork publication.
Add a new scanning framework or equivalent CI gate: unnecessary for retiring this identified auxiliary registration. A source check alone also does not stop an already-running workflow from publishing another snapshot.
Remove native platforms, Markdown regressions or assertions: changes required coverage rather than addressing resource ownership.

Consequences

Upstream PR updates no longer launch this registered screenshot workflow. Native tests still run normally, including for docs-only changes. Source-fork screenshot builds and artifacts remain available for review with their source identity.

At cleanup time only one of the identified PR screenshot entries remained. Removing it freed 1,351,294,893 bytes; the other previously observed entries were already absent and are not counted as deletions performed by this change. The resulting inventory retained all ten default-branch source/target snapshots, totaling 9,023,076,490 bytes. No cache quota or runner-plan setting changed.

This retires one upstream registration, not arbitrary future workflow files. New auxiliary workflows and any re-enablement still need resource/ownership review. Dependency or SDK changes can legitimately cause cold builds; cache retention does not establish a fixed CI duration.

Validation

The disable API returned 204; a fresh GET reported disabled_manually for upstream workflow 366570225 and active for fork workflow 366585525.
No screenshot run was active or queued, so no workflow run was cancelled.
Deleting cache 8179623547 returned 204. Readback showed the selected PR cache absent and every pre-cleanup default-branch cache ID still present.
The main-branch rules read before and after the operation were identical.
The preceding main native run 36362067181 completed successfully and repopulated the shared snapshots. A subsequent PR run measures reuse; this operation does not claim that the eight-minute target has already been achieved.

Supersedes

None. The coverage and cache boundaries in shared native PR caches and shared macOS runners remain unchanged.

Revisit when

An upstream-owned screenshot workflow has an agreed opt-in trigger, source identity and bounded read-only cache use. Re-enabling the recorded registration requires an explicit maintainer action after that review; a documentation edit is not an enablement operation.

6.1 KiB Raw Permalink Blame History