[codex] speed up non-PnP resolve

[hardfist]

Summary

• use an exact-path cache for non-PnP resolver instances to avoid repeated semantic path hashing
• bypass tokio::fs blocking handoff on current-thread runtimes while keeping multithreaded behavior unchanged

Why

The single-thread resolver benchmark was spending a measurable amount of time in path-cache hashing and Tokio fs handoff overhead. These changes keep PnP-enabled behavior on the existing path while optimizing the non-PnP hot path used by the benchmark.

Impact

• faster non-PnP module resolution in the current-thread benchmark path
• no intended behavior change for PnP-enabled resolution

Validation

• cargo clippy --all-features -- -D warnings
• cargo fmt --check
• cargo test still has the same 6 pre-existing PnP failures observed on origin/main
• CodSpeed simulation baseline: resolver[single-thread] 54.77 ms, resolver[[single-threaded]resolve with many extensions] 132.01 ms
• CodSpeed simulation final: resolver[single-thread] 39.85 ms (-27.24%), resolver[[single-threaded]resolve with many extensions] 124.84 ms (-5.43%)

Artifacts

• baseline: optimization-artifacts/codspeed/20260518-223013-baseline2/run.log
• final: optimization-artifacts/codspeed/20260518-223853-exact-cache-plus-direct-fs/run.log

[codspeed-hq]

Merging this PR will improve performance by 7.71%

⚡ 5 improved benchmarks
❌ 2 regressed benchmarks
✅ 5 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Memory	resolver[[single-threaded]resolve with many extensions]	12.8 MB	13.4 MB	-4.48%
❌	Memory	resolver[resolve from symlinks]	12 MB	12.9 MB	-7.09%
⚡	Simulation	resolver[multi-thread]	62.6 ms	60.7 ms	+3.04%
⚡	Simulation	resolver[[single-threaded]resolve with many extensions]	143.5 ms	133.1 ms	+7.83%
⚡	Simulation	resolver[resolve from symlinks multi thread]	114.9 ms	103.1 ms	+11.41%
⚡	Simulation	resolver[resolve from symlinks]	186.2 ms	167.6 ms	+11.08%
⚡	Simulation	resolver[single-thread]	55.9 ms	40.6 ms	+37.85%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf/micro-opt-resolve (1dd4c48) with main (0616bdc)}