[jlongster]

Summary

• Remove cache-on-failure: true from the Swatinem/rust-cache step in the Windows workflow, so it falls back to the action's default (false) and only successful jobs save a cache.
• Why: on PR fix(daemon): survive thread spawn failures #29 the arm64 job hit the 30-minute timeout and was cancelled, but cache-on-failure still let it save a cache entry. That entry was about half the normal size (~226 MB vs ~523 MB). The re-run restored the partial cache and hung again. Deleting that entry fixed it. With this change, a failed, cancelled or timed-out run can no longer save a broken cache.
• Only the Windows workflow changes. core.yml uses rust-cache without cache-on-failure, and publish.yml doesn't use rust-cache, so neither needs a change.
• This is independent of the fix(daemon): block subscriptions instead of polling #28–fix(daemon): wait for readiness instead of polling #31 stack.

Validation

• ruby -ryaml -e 'YAML.load_file(".github/workflows/windows.yml")' parses the workflow.
• The Windows checks on this PR run with the updated cache settings.

[jlongster]

This was folded into anomalyco/opencode#59694. The ported Windows workflow no longer saves the cache from failed runs.