perf: cut the fixed ~100ms guest process-launch charge to ~36ms

[WyvernMonarch]

Summary

Every guest process launch (each vm.process.exec, i.e. every bash tool call in an exec-style workload) paid a fixed ~100 ms charge. Stage-by-stage instrumentation of the warm path (92% coverage of the measured p50) attributed most of it to two avoidable costs, fixed here:

• perf(wasm): stop copying the guest module through a JS Array. The memory-limit enforcement path copied the module bytes element-wise through a JS Array on every exec — ~29 ms per launch for the shipped coreutils payload. Replaced with subarray views + a single Buffer.concat; parsing is unchanged.
• perf(v8): memoize the snapshot cache key instead of re-digesting per exec. The snapshot cache key re-digested the bridge + userland bytes on every exec. The key is now memoized keyed by content (unit tests assert the memo agrees with a fresh digest and that the key is dep-keyed over both bridge and userland, so no cross-content reuse).

Measured effect

Same-session sandwich (base → fixed → base restored), release sidecar, darwin arm64:

• direct vm.process.exec (n=25): p50 98.95 → 36.51 ms (2.71×), p95 100.03 → 38.20; restored baseline 99.16 confirms attribution;
• through pi's shipped bash tool (n=20): p50 99.5 → 36.4 ms (2.73×); read/edit/write and local-bash controls unmoved;
• guest stage attribution: memory-limit enforcement 29 → 0 ms, wasi.start 22 → 17 ms.

Equivalence

enfml_equiv.mjs over the shipped payloads: 271 checks, 0 divergences across 120 real wasm modules (module parse results byte-equivalent pre/post). New unit tests: snapshot_cache_key_memo_agrees_with_a_fresh_digest, snapshot_cache_key_is_dep_keyed_over_bridge_and_userland. agentos-v8-runtime --lib: 156 passed.

Notes

• Measurements are darwin arm64; the ratios were stable across two sessions but Linux numbers were not taken.
• While profiling we observed AGENTOS_WASM_SNAPSHOT_RUNNER=off measuring faster than auto on the pre-fix baseline — unexplained and possibly worth its own look; this PR does not change that logic.
• The snapshot-key memo is a small FIFO; behavior under many distinct userland bundles or high VM concurrency was not stress-measured.

🤖 Generated with Claude Code