test(assets): move cache cold-vs-hit timing out of the gate into a bench

[guysenpai]

Why

tests/assets/cache_diff.zig (M0.6 / E4) asserted wall-clock ratios inside the zig build test correctness gate — hit < 10 ms (absolute) plus miss > hit*20 (relative). Both are fragile across hosts and CPU load: a single cache-hit sample can spike on a page fault / AV scan / cold directory, which red-fails on slower or Windows CI runners. The spike also inflates the relative threshold (e.g. a 3 ms hit ⇒ needs miss > 60 ms, but a ReleaseSafe cold cook is only ~13 ms ⇒ fail).

This was flagged as pre-existing M0.6 debt during M0.7 / E3 Windows validation — see briefs/M0.7-ipc-scm-rights-windows-fuzz.md § Acted deviations → "Known debt left untouched" — to be resolved separately by making the assertion tolerant or moving it out of the gate.

Measured evidence (idle dev Mac, ReleaseSafe, 16 MiB asset): cold p50 12.7 ms, hit p50 184 µs (69× speedup) — but hit max 2.98 ms. On a loaded Windows runner that single-sample spike breaks both the old absolute and relative checks.

What changed

Picked option 2 — move the timing to the bench suite, keep a deterministic functional assertion in the gate. This is the only option that makes the gate fully deterministic + cross-host, and it matches the repo's established methodology: every perf number lives in a separate bench-* step (ECS, IPC RTT, adler32, paeth, etch), archived and non-blocking, measured under the opposable protocol on the reference machine — never inside zig build test.

• tests/assets/cache_diff.zig — stripped all wall-clock; the gate now asserts only deterministic facts: the miss → hit transition and a byte-identical cached artifact. Asset shrunk 16 MiB → 256 KiB (the large size only existed to make the cold cook expensive for timing, now the bench's job), keeping the gate fast on every host.
• bench/asset_cache.zig (new) — zig build bench-asset-cache, multi-sample cold-vs-hit differential, --smoke for CI, writes the gitignored bench/out/asset_cache_<os>.md. Mirrors the adler32 / paeth / render bench conventions.
• build.zig — registered the bench-asset-cache step alongside the other M0.6 benches.

Verification

Gate	Result
zig build test (Debug)	exit 0
zig build test -Doptimize=ReleaseSafe	exit 0
zig build lint / zig build	exit 0 / exit 0
zig fmt --check	clean
pre-push hook (build + test + test-release)	all ✔️

(The macOS full-suite failed command lines for events/plugin_loader are pre-existing spurious output; the suites exit 0.)

Scope

No IPC / M0.7 code touched. Branched off main (this is an independent M0.6-debt fix, not stacked on the unmerged M0.7 work).

🤖 Generated with Claude Code