#205 follow-up: Phase 7 warm-cache resolve bench (resolve_warm) #215

Context

crates/fbuild-library-select/benches/resolve_cold.rs measures the cold path of resolve(...) (P-02: ≤ 200 ms cold for a typical project). Phase 4 (PR #212) shipped resolve_cached(...) with a KvStore-backed memo. There is no companion warm bench yet, so the AC#5 / P-01 goal ("warm library-selection ≤ current fbuild + 50 ms") has no measurement.

bench/fastled-examples/ is reserved for the real-FastLED matrix once ~/dev/fastled is wired in; that is a separate, larger follow-up. This issue covers the per-crate criterion bench that runs against the existing MiniFramework and exercises the cache-hit path — same fixture as resolve_cold, but pre-populated KvStore and asserts the second call returns a cache hit.

Scope

1. Add crates/fbuild-library-select/benches/resolve_warm.rs:
   • Reuse MiniFramework from fbuild-test-support (same as resolve_cold).
   • Call resolve_cached once outside the timed loop to populate the cache.
   • Bench measures only the second resolve_cached invocation (the hit path).
2. Add the [[bench]] entry to crates/fbuild-library-select/Cargo.toml.
3. Wire it into .github/workflows/bench-205.yml matrix as a new lane with a target threshold (e.g. P-01-mini: ≤ 5 ms warm against MiniFramework — actual threshold can be tuned after first run).
4. Update crates/fbuild-library-select/benches/README.md describing both benches.

Out of scope

• The bench/fastled-examples/ real-matrix harness (gated on access to a checked-out FastLED tree).
• Wiring orchestrators to use resolve_cached (separate follow-up — required before any non-synthetic warm bench can mean anything).

Acceptance

• uv run soldr cargo bench -p fbuild-library-select --bench resolve_warm runs to completion locally.
• Bench reports a warm-path number distinctly faster than cold.
• CI workflow runs both benches and surfaces the numbers as artifacts.

Refs: #205 Phase 7 (P-01 mini-bench).