sparkles:fuzzy benchmark baseline

Recorded 2026-08-14 from the F0 working tree based on e0df6b19.

This is the first lifecycle-aware local baseline, not a portable performance claim. Run it with:

dub test :fuzzy -b bench -- --bench

The benchmark build uses release mode, optimization, inlining, -O3, and -mcpu=native. This run used LDC 1.41.0 (DMD 2.111.0, LLVM 18.1.8), Linux 6.18.26, and an AMD Ryzen 9 7940HX. Frequency scaling and normal workstation noise were enabled.

Benchmark	Construction boundary	Median
parse + analyze + glob compile	included; interactive code-path query	3.5 µs
glob execution	compiled program excluded; one path	2.7 µs
match score + positions	parsed query/workspace construction excluded; one path	3.8 µs
top-K generation	reset included; 64 offers, 20 retained/paged	1.2 µs
one-candidate generation	parse, accumulator reset, search, rank, and page included	6.1 µs

Each row alternates an input or iteration order at runtime and asserts a valid result, preventing the optimizer from timing only a constant validation path. The benchmark metadata printed by the runner records profile, tier, corpus, and construction for comparisons.

Comparison policy

Store future snapshots with compiler, CPU, ISA, build flags, corpus identity, and sample distribution. A designated-runner regression is actionable only when it exceeds 5% and a 95% confidence interval excludes zero for wall time, retired instructions, or cache misses. A -mcpu=native snapshot never gates a different ISA.

An fzf-algorithm C comparison is informational and valid only for the shared ASCII, maxTypos = 0, score-only workload with equal construction and output work. Ranking-quality gates remain exact D-side admitted sets, scores, positions, pages, and generation state transitions; different acceptance semantics cannot be called equal work.