Remaining low-ROI-but-real allocation nits after the ~−75% parse+build campaign (2026-07-02 audit + perf effort):
- operator-token interning — multi-char merged operators aren't interned (symbols currently only via a static single-char ASCII table).
- Semantics-phase allocation benchmark — the campaign gated Parse+Model; the Semantics/binding phase has no bytes/op gate yet.
- ToLowerInvariant dedupe — collapse residual duplicate
ToLowerInvariant call sites.
Rule: benchmark-gated — bytes/op is the judge (bench -- alloc / PipelineBenchmarks), no perf change ships without a measured delta; update docs/parser-performance.md on any measured gain (repo policy).
Acceptance: each item measured A/B; land only net wins; dashboard + the parser-perf memory updated.
Remaining low-ROI-but-real allocation nits after the ~−75% parse+build campaign (2026-07-02 audit + perf effort):
ToLowerInvariantcall sites.Rule: benchmark-gated — bytes/op is the judge (
bench -- alloc/PipelineBenchmarks), no perf change ships without a measured delta; updatedocs/parser-performance.mdon any measured gain (repo policy).Acceptance: each item measured A/B; land only net wins; dashboard + the parser-perf memory updated.