RUE-1223: make compiler timing contention-free

[steveklabnik]

Summary

• replace the compiler timing collector's per-transition shared lock with thread-local accumulation and bounded publication
• preserve exact nanosecond phase accounting with independent root-union and phase-band reducers
• add compiler-owned body-closure, CFG/optimization, and backend boundaries without teaching the generic query runtime about compiler phases
• cover real registered-batch publication, cancellation, deterministic merge, corruption detection, and phase attribution
• document the collector contract and the complete local performance ledger

Why

Timing collection ran on parallel query hot paths and serialized every span close and phase transition through one shared mutex. This change keeps the hot path local and merges once at worker completion or report generation, so instrumentation can coexist with compiler parallelism.

Performance

Alternating release-build pairs on an Apple M5 measured median paired instrumentation overhead of:

• Meridian: +1.65% (5 pairs)
• Caldera: +0.87% (11 pairs)

The ratio of Caldera's unpaired marginal medians was +2.64% on the noisy interactive host; raw samples and the pairing rationale are preserved in the acceptance ledger. Startup medians were both 0.01 s at the host timer's resolution.

Unattributed compiler-root time fell from 39.0% to 12.0% on Meridian and from 53.1% to 22.3% on Caldera. Instrumented and uninstrumented output hashes matched for both workloads, and the phase partition balanced exactly in integer nanoseconds.

Validation

• scripts/rue fmt
• scripts/rue quick (30/30 targets)
• /opt/homebrew/bin/bash clippy.sh (63 targets)
• RUE_TEST_TIER=premerge ./test.sh (78/78 targets)
• release Meridian and Caldera output-parity and exact-accounting smokes
• independent high-effort adversarial review: no remaining blockers

Fixes RUE-1223.