#205 follow-up: Phase 7 FastLED-examples warm-cache bench harness

[zackees]

Context

crates/fbuild-library-select/benches/resolve_warm.rs (#216) measures the cache-hit path against the synthetic MiniFramework fixture (~3× speedup, ~1.7 ms warm). That's a per-crate sanity guard, not the AC#5 measurement: the issue's perf criterion calls for the warm path to add ≤ 50 ms over current fbuild on the FastLED examples matrix.

bench/fastled-examples/ was scaffolded as a placeholder in #211 with a README that says it waits on Phase 4 + a follow-up to land the harness. Phase 4 is in production now (#212 + #217), so this is unblocked.

Scope

1. Add a Rust binary under bench/fastled-examples/ (or cleaner: a workspace-member crate fbuild-bench-fastled-examples) that:

   • Discovers FastLED examples from a configurable path (env FASTLED_DIR default ~/dev/fastled).
   • Picks a representative subset (e.g. Blink, Pacifica, Animartrix, Audio, BlurBenchmark, Chromancer — 5–10 sketches that span the simple/complex spectrum).
   • For each example, builds a synthetic project: the .ino as src/main.cpp, FastLED's own src/ as a project library, plus a Teensyduino-class framework library set (start with MiniFramework-extended; can grow to a real Teensyduino tree if needed).
   • Runs fbuild_library_select::cache::resolve_cached cold and warm against each (example, lib-set) pair.
   • Prints a table: example | cold (ms) | warm (ms) | hit-confirmed.
   • Asserts every warm call returns from_cache = true (catches silent re-misses).
   • Optional: emit report.json for future PR-comment diffing.

2. CI wiring in .github/workflows/bench-205.yml:

   • Add a workflow_dispatch-only lane (no automatic PR trigger — needs ~/dev/fastled checkout step).
   • The lane checks out FastLED at a pinned ref, runs the harness, uploads the report as an artifact.

3. README updates:

   • bench/fastled-examples/README.md — replace the "waits on Phase 4" framing with actual usage instructions.
   • bench/README.md — list the new harness.

Out of scope

• Per-board ELF-size acceptance gates — those live in acceptance-205.yml and tests/teensylc_acceptance.rs etc.
• Driving the full fbuild build pipeline — this bench measures only the resolver, not the compile/link chain.
• Threshold-based CI gating — first capture a baseline, then a follow-up adds the ≤ +50 ms gate.

Acceptance

• uv run soldr cargo run -p fbuild-bench-fastled-examples (or equivalent) runs to completion locally with FASTLED_DIR pointing at ~/dev/fastled.
• Report shows distinct cold vs. warm timings per example.
• All warm calls confirm from_cache = true.
• clippy --workspace --all-targets -D warnings clean.

Refs: #205 Phase 7 (AC#5).