Generate dist in CI and validate against generated output

[KaiKitJeffreyChan]

Summary

CI currently validates dist/ files committed to git. Once dist is published as release assets (#957), CI should generate dist/ itself and validate the generated output instead. Part of shop/issues-taxonomy#544.

What changed

• tests.yml: adds an explicit Generate distribution files step to the test_integration and vet_schema matrix jobs. The step removes the checked-out dist/ tree (so checks can never read committed or stale files) and regenerates en from data/ and the current VERSION. Unit-test jobs skip it. Generated files are ephemeral job inputs — no flattening, gzip, or release writes (publication stays in the stable/unstable asset workflows).
• Removes distribution_matches_data_test.rb: comparing regenerated output against a committed dist/ baseline is obsolete once generated files are no longer versioned.
• Removes the /generate_dist commit-back workflow: superseded by generation-in-CI. It was already broken — it invokes make --file=Makefile build, and the Makefile was deleted in 354723b.

Gitignoring dist/ and rejecting PRs that touch it is deliberately not part of this PR — that's the final cutover task, which can't land while dist/ is still committed.

Validation

Full workflow run on a fork: all 4 jobs green.

• test_integration / vet_schema logs show rm -rf ../dist → generation from VERSION → checks passing against the fresh tree (11 runs, 527791 assertions, 0 failures).
• test_unit skips the generation step.

Known gaps / follow-ups

1. Committed dist/ can drift until the cutover (shop/issues-taxonomy#546). With distribution_matches_data_test removed, no check compares the committed dist/ against data/ — a data PR that skips regeneration merges green and main's dist/ stays stale until the next release (GenerateReleaseCommand regenerates all locales). This only affects consumers reading dist/ from main rather than from releases. Open question for reviewers: is main-branch dist/ a consumption surface we support during the interim? If yes, I'll re-add the deleted test as a non-blocking job (continue-on-error: true, same shape as orphan_detection) until Remove old tooling #546 lands.
2. MANIFEST_KEY repo secret is now orphaned. The removed workflow was its only consumer (it keyed-BLAKE3-signed the artifact handed between its own two jobs). Deleting the secret from repo settings loses nothing — will remove it after this merges.
3. Cutover prerequisite logged on Remove old tooling #546: generate_release_command.rb:81 version-badges dist/README.md via File.read. When dist/ is deleted/gitignored at cutover, that line must be removed in the same change or the release command crashes with Errno::ENOENT.
4. Post-merge verification: the Add "Other" entry as default base color to Color attribute #544 acceptance criterion ("representative data/ change runs green through the new CI") is demonstrated on a fork above but not yet on this repo — the first data/-touching PR after merge confirms it.