v0.19.2 — finish the binary-heavy directory slowdowns

A follow-up to 0.19.1. That release stopped one code path from reading
unsupported binaries as text; two others still did. On a 102-file, 8.6 GB
GeoTIFF/shapefile directory:

• scan_directory: 152s → 0.04s. The code-health pass was read_text()'ing
  and splitting every file — including multi-GB rasters — to count
  definitions and references. It now skips unsupported file-info stubs.
• preview_directory: 6.6s → 0.09s. The code map read each file in full via
  the generic-analyzer fallback before deciding whether to analyze it. It
  now checks that first and sniffs the first 8 KB for a NUL byte, so binary
  files are skipped without a full read. Text geodata (.sos, .geojson) is
  still read normally.

Also fixed: 0.19.1 added a bare "cache" to the directory skip list, which
matched on base name only and so silently pruned real source modules named
cache/ (e.g. src/cache/). Only dot-prefixed tool caches (.cache, .turbo,
…) are skipped now.

949 tests pass; golden output unchanged.