[Instructions -0.09% vs main; paired PGO estimate] Count captured token trees without intermediate parser steps

[xmakro]

Count the remaining token tree with a small explicit stack and advance the cursor directly. Keep the ordinary parser path for dummy closing spans and preserve delimiter counting and capture bookkeeping.

Comparison	Wall time	Instructions	Peak RSS
Complete PR vs main	-0.101%	-0.091%	-0.124%

Profile, complete PR vs main	Wall time
check	-0.413%
debug	-0.099%
opt	+0.210%

Negative percentages mean less elapsed time or fewer resources. The title is an individual local instruction-count estimate: instructions:u as an equal-case geometric mean over the 249 Primary cases, the metric the upstream perf bot summarizes; small changes may be measurement noise.

The two recorded repetitions give aggregate instruction estimates of -0.091% and -0.091% for the title comparison. Mixed directions are marked in the title. This repeatability check is not a confidence interval.

Both compilers use the same source base, CI LLVM, fixed standard libraries, ThinLTO and jemalloc. Each has separate nine-workload Rust PGO training. The stock collector measures all 20 Primary benchmarks: 249 Check/Debug/Opt and full/incremental cases, two observations per compiler per case, alternating baseline/candidate order by crate. No other campaign builds run during measurement. Percentages are equal-case geometric means. LLVM retraining and BOLT are omitted.

This branch has one commit directly on main. Its exact measured source passed the full UI and incremental suites plus the additional suites recorded in the build provenance.

Fixed main commit: da47efd27287edb02eed0b5a178a165c8f258e37. Measured source patch SHA-256: 88fd9ec69f74792c7540f8dcaf842c2d3f7c74e3206e338803d0d4c374702bb7.

Local raw measurements and per-case CSVs: independent-pr121-main-v1-comparison.csv in the September 6 investigation directory. Build/test provenance: pgo/independent-pr121-source-timing-v1/provenance.json.