t27b: native arm64 CI per PR on ubuntu-24.04-arm and macos-14 (off the Mac, off qemu)

Problem

No CI job builds or runs t27b. Every lane measures on the owner's Mac (load average about 180, three cargo builds at once in skill row Q45) or waits up to 40 min for the Railway lab, which runs master only and never a PR branch. The lab runs t27b under qemu-user on x86_64, so the native paths are never exercised there: macOS MAP_JIT/W^X, icache flush, the Mach-O emitter, and real FP and memory behaviour. cg_clif dropped qemu for native runners for the same reason.

GitHub gives public repositories free arm64 runners: ubuntu-24.04-arm, and macos-14 for Apple Silicon.

Change

A workflow t27b-native.yml, triggered by pull requests touching cli/t27b/**, bootstrap/src/** or specs/**:

ubuntu-24.04-arm: cargo test --release -p t27b with T27B_DIFF_SEED set to the run number (a new fuzz seed per run; failing seeds are printed), then t27b corpus specs --json natively, with --reference-cache restored from the lab's last cache artifact. The result is diffed against the ledger with python3 scripts/tri_loop/t27b.py ratchet.
macos-14: cargo test --release -p t27b only (MAP_JIT and Mach-O).
The check is named so that tri t27b ready can require it.

Done when

A t27b PR shows a native per-PR counted number and ratchet verdict within about 15 min, without the Mac or the lab.
tri t27b ready reads the new check, and the skill's local-build cap (Q45) can be relaxed.

🤖 Generated with Claude Code