ci: skip redundant diffr release builds and empty MIME tests

[sidkmenon]

Release crate verification currently waits for all five platform builds and Homebrew checks. On tag releases, cargo publish then verifies the packages again. Run the first verification alongside platform builds and skip the duplicate compilation during publication.

After the full cargo publish --locked --dry-run succeeds, save SHA-256 checksums of its three .crate archives. The publication job uses the same commit and Cargo 1.99.0, recreates the packages without compilation, and must match those checksums before publishing with --no-verify. The handoff is a checksum file from the same workflow run; no build-directory transfer or new persistent cache is needed. Manual runs exercise this handoff without publishing.

All platform builds, archive smoke tests, Homebrew checks, and the release approval gate remain. Verification failures or checksum mismatches block publication. Cargo installation emits timing reports as separate artifacts, and checkouts do not persist credentials. Grammar optimization and additional caching are deferred. Remove the obsolete MIME job from PR and main CI: it runs cargo test -- --ignored, but there are no ignored Rust tests left. A completed job spent 2m13s compiling and then ran zero tests. The ordinary test suite remains unchanged.

Validation

• Actionlint, Rust formatting, diff checks, and both existing xtask tests passed locally.
• Cargo 1.99 packaging, multi-package publication dry runs, and individual core/CLI publication dry runs produced byte-identical archives locally. These probes used --no-verify and did not publish anything.
• A checksum probe accepted the unchanged archive and rejected a modified archive. Publication-gate checks covered successful, failed, cancelled, and skipped prerequisites; manual runs cannot publish.
• GitHub checks and manual release validation are running on da8103888; the subsequent commit only removes the empty MIME job and does not change the release workflow. No new end-to-end speedup is claimed until that run completes.

Baseline: the completed release validation spent 15m31s building/packaging Intel Mac, then 14m52s verifying crates. Parallelizing these removes the serial verification delay; the exact improvement depends on runner and queue times. Skipping the later publication rebuild is a separate saving on actual tag releases.