Speed up path matching and add label-only traversal

[andrew]

Reduce path-matching overhead by storing pointers to corpus rules and rejecting invalid prefix boundaries before case folding. On Go 1.27.1, darwin/arm64, the median million-path benchmark fell from 274 ms to 209 ms across three runs. Matching the brief path inventory fell from 228 to 177 ns/path, with zero allocations. Corpus labels and evidence ordering are unchanged.

Add WalkMatch and Classifier.WalkMatch for visitors that only need a compact Set, plus CLI -labels-only. Both walk modes share traversal, limits and pruning semantics. For 10,000 files in a deep vendor/source/test layout, label-only traversal used 6.65 MB versus 34.05 MB for owned evidence results. Wide and monorepo disk timings were similar despite lower allocation, so this does not imply a speedup for every filesystem scan.

An isolated brief integration called Match on accepted paths inside its existing indexing loop, without adding file reads or another traversal. Six alternating pairs of full Engine.Run measurements over 10,000 synthetic files showed median scan overhead of 0.37%, with individual pairs between -0.37% and +1.03%. Import initialization was measured separately at roughly 1 ms. This experiment used minimal aggregation; storing richer per-file results would have additional costs. No brief integration is shipped here.

Update API documentation and CLI examples, add concurrent matching benchmarks, and cover label-only traversal through public APIs and the CLI, including vendor-root context, pruning, limits and errors.