perf(parse): give each thread its own regex cache pool

[jqnatividad]

Summary

regex! returned one global Regex per pattern. regex-automata gives only the first thread that uses a Regex the lock-free cache fast path; every other thread goes through mutex-protected cache stacks (util/pool.rs, MAX_POOL_STACKS = 8). Under a parallel caller like qsv stats --infer-dates, that contention ate ~10% of busy CPU (pthread_mutex_*, Pool::get_slow, Pool::put_value in a samply profile), and date inference cost +1.41 s CPU on 16 threads vs +0.75 s single-threaded.

regex! now returns a &'static LocalKey<Regex> holding a per-thread clone of the once-compiled regex. Regex::clone shares the compiled program (Arc) but builds a fresh cache pool (meta/regex.rs impl Clone for Regex), so each thread owns its pool. It's thread-local rather than a leaked &'static clone, so long-lived processes that churn threads don't accumulate lazy-DFA caches. The 23 call sites change mechanically: re.is_match(input) becomes re.with(|r| r.is_match(input)), and likewise for find/captures, whose results borrow only the haystack.

Benchmarks

qsv stats on NYC 311 (1M rows, 539 MB), 16 jobs, M4 Max, release build. Outputs are byte-identical before/after across --infer-dates, --everything --infer-dates, --prefer-dmy --jobs 1 and --dates-whitelist all:

run	before	after
--infer-dates	266 ms	225 ms (1.18×)
--infer-dates --dates-whitelist all	193 ms	155 ms (1.25×)
--jobs 1 --infer-dates	1.70 s	1.71 s (unchanged)

Testing

• cargo test, cargo fmt --check, cargo clippy --workspace --tests --all-features -D warnings: all pass
• qsv's stats/schema/apply/validate/moarstats tests (1,040) pass with this branch patched in

🤖 Generated with Claude Code