perf: the discovery sweep (~6.4s stat/fingerprint pass over the full tree) dominates every dated command

[iamtoruk]

Successor to #1034/#1057 (closed 2026-08-22). Measured on a real 21k-file / 9-provider corpus: warm, day-scoped codeburn today spends ~6.4s wall / ~5.6s CPU, and the dominant term is the scope-independent discovery pass — walking and stat()/fingerprinting every file in every provider tree regardless of the query's date range. Eliminating 89% of the redundant re-parses (#1057's mechanism, verified working) moved neither wall-clock nor CPU, proving the sweep owns the runtime.

Directions worth exploring (each needs its own never-lose analysis):

• Scope the WALK, not just the load: prune directory subtrees by mtime (a project dir untouched since before the range can't contain in-range appends — but beware: mtime of a dir doesn't propagate from deep children on all filesystems)
• fs-events/watch journal between runs (the resident serve child could maintain a dirty-set so warm queries stat only changed files)
• Batch stat via readdir-with-types + only fingerprint files whose (size,mtime) moved vs the cached envelope — measure how much of the 6.4s is stat syscalls vs fingerprint hashing vs JS overhead first

Baseline numbers and methodology live in the #1057 settlement comment. The prize: warm dated commands in low hundreds of ms on corpora this size — the 'sub-second warm' target the perf sprint originally aimed at.