🚧 fix: skip Jekyll pages build and prune superseded patch bundles on release

[jeswr]

🚧 Release-flow fix for the failing pages deployment

Note: This PR was generated by an agent on @jeswr's behalf, following up on the triage in #1845. Please review accordingly.

References #1845 (intentionally not "Closes" — the pages branch is still ~7.4GB until the separate one-off prune below is approved and executed, so the deployment failure can recur in the meantime).

This PR performs no pages-branch surgery. It only changes the release flow going forward.

What this changes

1. Skip the Jekyll build (.nojekyll) — immediate mitigation

The failing step is GitHub's implicit pages-build-deployment Jekyll build, which copies the whole ~7.4GB site on the runner and dies with No space left on device. scripts/post-webpack.ts now writes an empty .nojekyll marker at the bundle root, and the @qiwi/semantic-release-gh-pages-plugin config gains "dotfiles": true — without it the plugin's gh-pages globbing (**/* with dot: false by default) silently drops the file. Once the next release publishes it, Jekyll is skipped entirely. (The 1GB GitHub Pages site limit still applies, hence 2. and the one-off prune.)

2. Stop the regrowth — bound per-minor storage

Each release adds ~32MB (patch dir + latest + vMajor/latest + vMajor/vMinor/latest, each holding index.js + dynamic-import.js at ~4MB) and "add": true means nothing is ever removed. Now, after semantic-release publishes x.y.z, a new workflow step prunes the superseded patch dirs of minor x.y from the pages branch:

• Never touched: latest/, vMajor/latest, vMajor/vMinor/latest (all documented URL shapes in the README keep working), example/, other minors/majors, and dev/bench (written by the benchmark workflow — the prune only ever deletes numeric dirs directly under <major>/<minor>/, so it cannot reach it by construction).
• Removed: numeric patch dirs of the released minor other than the one just published — e.g. publishing 4.10.3 removes 4/10/1 and 4/10/2. Per-minor storage is thus bounded at one patch dir + the latest copies. (Pinning an exact older patch remains possible via the npm package; README updated to say so.)
• Mechanics: blobless sparse clone (--filter=blob:none + cone sparse-checkout of <major>/<minor> only), so the runner never materialises the large branch; ts-node scripts/prune-pages decides what to delete; push uses the same GH_TOKEN as the release (a github.token push would not trigger the Pages deployment) with a fetch+rebase retry loop in case the benchmark writer races.

The prune logic lives in scripts/prune-pages.ts (exported + CLI with --dry-run) and is covered by a new jest suite.

Dry-run evidence (local, no publish, no pages push)

Fixture tree mirroring the pages branch (latest/, 4/latest, 4/10/{1,2,3,latest}, 4/9/{9,latest}, 3/22/{0,latest}, 3/latest, example/, dev/bench/, .nojekyll):

$ npm run pages:prune -- --name=v4.10.3 --root=<fixture> --dry-run
Would remove 2 superseded patch dir(s) for v4.10.3:
  - 4/10/1
  - 4/10/2
Kept within this minor: 4/10/3, 4/10/latest
# tree verified byte-identical after dry run

Real run against the fixture removes exactly 4/10/1 and 4/10/2; everything else (including dev/bench and all latest shapes) verified untouched.

End-to-end rehearsal of the exact workflow step against a local bare repo standing in for origin (blobless sparse clone → prune → commit → push): resulting pages tree keeps .nojekyll, 3/22/{0,latest}, 4/9/{9,latest}, 4/10/{3,latest}, all latest shapes, example/, dev/bench — only 4/10/1 and 4/10/2 dropped. A second rehearsal with a competing dev/bench commit pushed mid-flight confirmed the retry loop recovers (rejected push → fetch+rebase → push OK on attempt 2) with both the bench update and the prune preserved.

.nojekyll path: post-webpack.ts smoke-tested in an isolated fake repo — bundle/.nojekyll (0 bytes) created alongside the usual dynamic-import.js and the three latest copies.

Validation: eslint on touched files exit 0; tsc --noEmit on the new/changed scripts + test exit 0; new jest suite 6/6 passing; workflow YAML parses.

One-off prune (separate sign-off)

Not part of this PR — awaiting explicit approval from @jeswr before anything is done to the pages branch. This PR only bounds future growth; the branch is already ~7.4GB (2,974 files, 188 minor dirs), ~7x the 1GB Pages limit. The proposed one-off action is:

• Rewrite the pages branch as a single orphan commit (shrinking clone/checkout size, not just the deployed tree) containing only: dev/bench (benchmark history), example/, latest/, every vMajor/latest, every vMajor/vMinor/latest, and the current patch dir of the newest minor, plus .nojekyll.
• Consequences: historical per-patch URLs (/x/y/z/index.js for superseded patches) stop resolving — anyone pinning them should move to vMajor/vMinor/latest or the npm package, where every published bundle remains available forever (dist/ ships in the eyereasoner package). All documented URL shapes in the README keep working.
• Force-push of the rewritten branch is the only step that needs the sign-off; it is independent of (but complementary to) this PR and can happen before or after merge.

Happy to adjust retention (e.g. keep the last N patch dirs per minor instead of one) if preferred — the prune script takes the decision in one place.

🤖 Generated with Claude Code