[helly25]

Ranked fuzzy matching now reuses scoring memory and caches pattern normalization instead of recomputing it per file. Paired 100,000-file measurements improve about 4-9% on macOS and 1-6% on Linux, depending on input and worker count. Valid matches after long gaps also retain their zero-quality result instead of disappearing.

AG;DR

• O03 from the reconciled docs/performance-analysis.md / TODO.md inventory; follows merged Share collection name and root storage with owned paths #1041/Prepare simple native globs and avoid pattern copies #1042. No outstanding PR dependency.
• Compute identical-pattern normalization in linear time and cache it in immutable parsed Fzf terms. Preserve full dynamic programming against candidate input.
• Reuse two uninitialized score rows owned by the coordinator or each existing persistent matching worker. No new pools, threads, locks, refcounts, flags or exception cleanup. Initialize every first-row cell before any read.
• Allocate exactly the first input size; double later capacity up to 65,536 cells/row, 512 KiB retained total. Larger inputs use ephemeral storage. Moved-from workspaces remain reusable.
• Correct the inherited unmatched sentinel: a score of -1484 after 1,500 skipped bytes is valid, normalized to zero. Explicit first/later-row long-gap regressions supplement the general fuzz input cap.
• Preserve ranking, inversion, anchors, folding, short-circuit and deferred replay. Levenshtein/shingle model allocation is outside this change.

Evidence and tradeoffs:

• docs/benchmarks/fuzzy-score-workspace-20261102.md publishes all raw samples, scripts, source/binary identities, four-way causal controls and rejected initial-capacity trials in lossless compressed artifacts.
• Final reference 44381e2 has the same runtime tree as merged main 2b21790. LLVM 23 release + ThinLTO. Three sessions, interleaved nine rounds, fastest seven; independent exact native output oracle. Linux tmpfs with child affinity 1/3/10; macOS warm APFS, requested workers 1/3/10, no affinity.
• 4,320 final native invocations including warmups. Scored kernels improve 29-54%; 10k full-engine scored selection improves 4-11%. Whole-command gains are smaller because traversal and output remain.
• Retain slower controls: bare fuzzy kernel +3.02% macOS / +4.25% Linux; slowest paired native median +1.65% / +1.85%. Preparation is slightly slower. Do not claim every fuzzy model or invocation improves.
• Reject the 64-character first-allocation minimum after cross-platform capacity trials. Allocation reduction alone did not prove a speedup.
• Isolated ASan counters, not sanitizer timings: 1,000 fresh queries allocate 1,000 buffers / 64,000 bytes; warmed reused scratch allocates zero on both hosts. Build-output executable grows 624 bytes macOS / 8,488 bytes Linux, before release stripping/compression.

Validation:

• Complete release and extension suite: macOS 304/304, Linux 304/304 with working FUSE explicitly required.
• Seven targeted release targets and isolated allocation diagnostics pass on both hosts. Tests cover normalization against full alignment, length/case/separator variants, buffer growth/caps/moves, independent thread scratch and long gaps.
• Local scoped clang-tidy passes all six runtime/test translation units and normal commit hooks pass. Normal commit and push hooks pass. Linux MSan fuzzy tests and fuzz corpus also pass. Published-head CI results will be recorded here as they complete.

Published head 5ddc685e18ea9cbc19f38b6084c58ac55e4bb3d4 passed all required checks, including coverage, GCC, Linux/macOS tests, ASan, TSan, MSan, clang-tidy, fuzzing, and the final benchmark aggregate.