test(bench): add six performance benchmarks for core hot paths

[HanSur94]

Summary

Grows benchmark coverage across the FastSense performance surface, added incrementally via /bench-evolve (one focused benchmark per iteration). Additive only — new benchmark scripts under benchmarks/, no library or production code changed. Each benchmark is deterministic, labels the active MEX-vs-fallback path, and was verified in MATLAB R2025b (and Octave where the feature supports it).

Kernel microbenchmarks — every MEX kernel wired into production

• bench_downsample_kernels.m — LTTB + MinMax downsampling throughput (the core resolve→downsample→render kernels; previously LTTB was never timed). MEX: ~344 Mpts/s LTTB, ~1100 Mpts/s MinMax at 20M; O(N) confirmed.
• bench_violation_cull.m — fused threshold detect + pixel cull, exercising both the constant and time-varying (ZOH) threshold branches.
• bench_binary_search.m — per-call latency of the O(log N) viewport-clip / bucket-boundary search (the most ubiquitous hot-path call); ~1 µs/call, log-like scaling.

Composite-path benchmarks

• bench_fastsense_multiline.m — live-update refresh rate vs line count on one axes (564→39 Hz over 1→64 lines; sublinear per-line cost).
• bench_detached_mirror_refresh.m — refresh overhead of detached live mirrors, i.e. the project's headline constraint ("detached live-mirrored widgets must not degrade refresh rate"). Finds mirrors add real, increasing overhead (+35%→~+200% over 1→4 mirrors). MATLAB-only (Octave-skipped — see findings).
• bench_derived_resolve_chain.m — DerivedTag full-chain recompute cost vs dependency depth (cold ~linear in depth; warm cache flat ~0.02 ms, proving the lazy-memo).

Coverage note

benchmarks/.reports/coverage.md tracks what's covered, what remains, and the methodology lessons baked into the noisier composite benches (warmup, amortized vs per-tick median for bimodal drawnow('limitrate') timing, min-over-reps for GC-bursty recompute).

Findings surfaced (reported, not fixed here — additive-only remit)

1. Three MEX kernels are test-only — compute_violations_mex, to_step_function_mex, resolve_disk_mex ship with parity tests but have no production .m call site. Either dead code or missing wiring.
2. Octave-incompatible deserialization — DashboardWidgetRegistry.fromStruct (libs/Dashboard/DashboardWidgetRegistry.m:92) uses feval([className '.fromStruct'], s), which Octave can't resolve. This breaks serialized dashboard load and widget detach under Octave (a "fully supported" runtime). One-line str2func fix candidate noted in the coverage file.

Test plan

• All six run cleanly in MATLAB R2025b on Apple Silicon (real MEX kernels).
• Portability verified: bench_downsample_kernels, bench_violation_cull, bench_binary_search, bench_fastsense_multiline, bench_derived_resolve_chain run under Octave 11; bench_detached_mirror_refresh cleanly skips on Octave (guarded — the detach feature is currently MATLAB-only per finding Add datetime X-axis support #2).
• These are baseline-able by /bench-guard and /perf-watch; no production code touched, so no regression risk to the library.

🤖 Generated with Claude Code