perf: reduce compiler allocation overhead and close the JIT gap

[toakleaf]

Summary

• remove ineffective sync.Pool lookups from hot AST constructors and embed base nodes in high-volume AST types
• share immutable built-in function definitions while preserving per-compilation child registries
• replace the hottest parser regex probes with exact byte scanners
• replace recursive visitor pool traffic with visitor-local storage
• reuse the immutable default() guard sentinel
• add a benchmark for the exported Compile API and update the benchmark documentation
• document GOGC=200 as an opt-in throughput setting rather than changing the process-wide runtime from the library

Performance

Measured on Apple M1 Max, Go 1.24.5, Node.js 24.4.1, 212 benchmark files:

Benchmark	Before / JS	After / Go
Go BenchmarkLargeSuite median	143.10 ms	115.17 ms
Go allocations per build	~1.321M	~1.147M
Fresh-process full build	JS 207.08 ms	Go 117.46 ms
Cold per-file average	JS 799.19 µs	Go 505.55 µs
Warm isolated-file average	JS 402.75 µs	Go 476.68 µs
Warm aggregate median, default GC	JS 102.57 ms	Go 112.07 ms
Warm aggregate median, GOGC=200	JS 102.57 ms	Go 96.78 ms

The retained code changes improve the core Go suite by about 19.5%. Realistic fresh-process builds are 1.8x faster than Less.js. Peak-JIT isolated-file performance remains about 18% ahead of default Go, while GOGC=200 moves the aggregate warmed Go suite slightly ahead with a larger heap target.

GC trace sampling over 15 builds:

Setting	GC cycles	Maximum observed heap goal
GOGC=100	553	10 MB
GOGC=200	193	15 MB
GOGC=400	76	27 MB

Investigation notes

Several ideas were measured and reverted rather than included:

• explicit parse-tree release reduced allocations but slowed the suite by about 5.5%
• a generic no-capture regexp fast path regressed performance
• direct Unit allocation and combinator trimming changes had no reproducible effect

Validation

• go test ./...
• go test -race ./less -run='TestCompile|TestFactory|TestIntegrationSuite/main/functions|TestFunctionRegistry' -count=1
• go run scripts/test.go integration — repository baseline 316/317, one existing compile failure, zero output differences
• node scripts/bench-compare.js
• node scripts/bench-compare-suite.js