ci: merge bench-compile-gate + bench-report into single bench job

[dryotta]

What

Replaces bench-compile-gate + bench-report with a single bench job, and fixes the post-merge tree-hash skip for bench work.

Why

1. Compile work was duplicated per PR

The two jobs each ran cargo bench --profile release-ci under independent Swatinem prefix-keys (bench-compile-gate and bench-report), compiling the six [[bench]] targets twice per PR. The release-ci profile is fast (lto=off, codegen-units=16), but a cold compile is still ~8-12 min. Two independent caches also mean two independent cache-state miss windows.

2. Post-merge tree-hash skip was bypassed

ci-gate-inputs computes a skip boolean that is true when the post-squash tree on main is byte-identical to a PR head whose ci.yml run already concluded success. Every heavy job's if: carries code == 'true' && skip != 'true' except bench-compile-gate and bench-report, which only checked code == 'true'. Result: every push-to-main re-ran the bench compile even on tree-identical squash-merges.

How

Single bench job structure:

Step	continue-on-error	Notes
Setup (checkout, toolchain, cache, apt, stage CLI placeholder)
cargo bench --no-run --profile release-ci	none	Blocking compile gate. Failure here fails the job and the merge gate.
Generate bench fixtures	true
cargo bench ... --output-format bencher	true	Reuses target/release-ci binaries no recompile.
Write perf summary	true + if: always()
Upload bench output	true + if: always()

Step-level continue-on-error: true (instead of job-level) means a slow or budget-busting bench shows as a red step inside an otherwise-green job. This is strictly more visible than the prior job-level continue-on-error, which hid regressions behind a job-level "advisory" label.

test-gate's needs: list and result check were updated from bench-compile-gate to bench. No other workflow callers reference the old names.

Trade-offs considered

• Two jobs with needs: chain + shared prefix-key. Cleaner separation, but Swatinem only saves on main (save-if: github.ref == 'refs/heads/main'), so on PR runs bench-report would still do a full release-ci recompile defeating the goal. Flipping save-if: true would write to the shared cache on every PR run (size churn + eviction risk).

• Critical-path impact. Negligible. Both former bench jobs were strictly under the slowest of {build matrix, native-e2e, installer-e2e}, so neither was on the critical path. The merged job's compile-then-run sequence (~3 min compile warm + ~1 min bench + ~10 s summary) is still well under critical-path jobs.

Cache migration

The old bench-compile-gate and bench-report Swatinem caches are now orphaned and will LRU-evict naturally. The new prefix-key: bench starts cold for the first main-branch run.