t27b: bench blocks handled as the reference path handles them (first rejection in 35 specs) #6179

Refs #6063

t27b (cli/t27b, the native AArch64 backend) refuses every spec with a bench block. On the Railway t27b lab run of 2026-10-04 (https://t27b-lab-production.up.railway.app/latest.json), BenchBlock is the first rejecting construct in 35 spec files and occurs 967 times in all.

The reference path is t27c gen + zig test. t27b should treat a bench block exactly as that path does: read what t27c gen emits for one (for example t27c gen on any spec under specs/ that contains bench). If the reference does not run benches as tests, t27b compiles the body natively (so a body it cannot lower still rejects the file) and does not count it as a test. If the reference runs them, t27b runs them too.

The steward (not the bee) merges, measures the corpus on the lab and reports on #6063. CI runs the cargo tests; do not build on the owner's machine.

Boundary

• cli/t27b/src/lower.rs
• cli/t27b/src/main.rs
• cli/t27b/tests/differential.rs

User Scenarios

1. A spec whose only non-test item t27b rejected was a bench block now gets the same verdict from t27b test as from the reference path.
2. A bench block whose body uses a construct t27b cannot lower still makes t27b corpus --blockers name that construct, not BenchBlock.

Requirements

• FR-001 -- bench blocks are handled the way the reference path handles them (state which way in the PR body, with the t27c gen output that shows it).
• FR-002 -- no interpreter fallback and no silently skipped body: a body that cannot be lowered is a rejection.
• FR-003 -- a differential test in cli/t27b/tests/differential.rs named bench_... over a spec with a bench block.

Success Criteria

• grep -c "fn bench_" cli/t27b/tests/differential.rs prints at least 1
• grep -c "BenchBlock" cli/t27b/src/lower.rs prints at least 1
• grep -c "interpreter fallback" cli/t27b/src/lower.rs prints 0