## Why
`generated_files.yml` only runs on `pull_request`, so its `cache-to:
type=gha,mode=max` wrote buildkit blobs into per-PR scopes that other
PRs cannot read — every new PR cold-built the codegen image (235–365s in
11 of 17 runs over the past week vs ~65s warm), and ~6 GB of duplicate
blobs pushed the repo's Actions cache to 9.9 GB of the 10 GB limit,
evicting the Playwright and pnpm caches that #1538 relies on.
## What
Adds `codegen_image_cache.yml`, which builds the image on pushes to
`main` touching its actual inputs (`codegen.Dockerfile`,
`packages/connect-python/**`, or the workflow itself) and exports the
cache to main's scope, readable by all PRs; it also supports
`workflow_dispatch` for manual re-seeding. The PR-side build in
`generated_files.yml` keeps `cache-from` but drops `cache-to`. Merging
this PR triggers the first seed automatically, since the new workflow
file matches its own paths filter.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Mish Ushakov
423a1b7302