perf: consolidate Scanner::walk_files traversals honoring per-check budgets

[hyperpolymath]

Deferred from CodeRabbit T3 on #185 (trivial severity, scanner-cold path).

Consolidating the seven walks in aletheia/src/checks.rs without changing per-check budget semantics is correctly-scoped heavy-lift perf work, out of scope for the review it came from. Track it here.

Acceptance: fewer traversals, identical budget semantics, 68+47 suite green, clippy -D warnings clean.

[hyperpolymath]

Resolved in e477b400 (PR #228), merged 2026-09-24.

ScanSet consolidates the per-check walks into a single traversal — one set of extension buckets, filled once — and caches it (files: OnceCell<ScanSet>). walk( no longer appears in aletheia/src/checks.rs; walk_into_set is the only traversal. Per-check budget semantics are unchanged, with three unit tests covering exactly that:

• test_scan_set_cap_is_per_bucket
• test_scan_set_filtered_buckets_cap_independently
• test_scan_set_routes_each_extension_to_its_buckets

Measured: 7 walks to 1; 212 ms to 150 ms on a tree of 18,602 files across 673 directories; verdicts byte-identical before and after.

The repo's own gate is green on the merge commit: rust-ci passed all five jobs (build debug + release, full test suite, fmt --check, clippy -D warnings, zero-dependency check) on both the PR head and the merged main.

This issue stayed open only because the PR body read "Closes #186 and #197" — GitHub pairs the closing keyword with the first issue only, so "#197" was treated as a plain reference. Nothing was outstanding in the code.