Overview perf + caching, gray-matter fix, ferrosearch search engine - #21

One dense fixture vault exercising every code path — weighted TF sources,
heading dedup and corroboration, bigram suppression, folder-token
exclusion, all noise patterns, scaffold skipping, depth limits, malformed
frontmatter, and both collapse outcomes — snapshotted across three option
sets. Any behavior change in the pipeline must show up as a snapshot diff.

Three behavior-preserving changes to the term pipeline:
- terms are tokenized once per text; bigrams derive from the token array
  instead of re-running the noise-stripping regex chain
- each noise regex runs only when a cheap necessary condition holds
  (an email needs @, a URL needs http, ...)
- occurrences are counted into a small per-text map and merged into the
  folder accumulators by unique term, preserving first-occurrence
  insertion order exactly (keyword tie-breaking depends on it)

Equivalence is enforced, not assumed: the pre-refactor implementation is
kept verbatim as an oracle, and a differential suite fuzzes both across
30,000+ adversarial strings and 60 randomized vaults, asserting
byte-identical output including Map-order-dependent tie-breaks.

3.7s -> 1.0s on a 5,000-note vault.

Single-entry cache in .napkin/overview-cache.json keyed by the whole-vault
fingerprint (paths + mtimes, shared with the search cache) plus the
resolved options. A hit costs one stat pass instead of reading and
tokenizing every note: ~25ms vs ~1s on a 5,000-note vault. Any file add,
remove, or touch invalidates; corrupt cache files are ignored and rebuilt.
Tests observe caching strictly through getOverview behavior, including a
frozen-mtime rewrite that must be invisible.

…gine

@shift-labs/ferrosearch 0.1.1 replaces minisearch in the search core. The
serialized index format is interchangeable, so existing search-cache.json
blobs keep working — pinned by a migration test asserting identical files
and scores from a minisearch-written cache. The cache write moves to
toJsonString() (ferrosearch has no toJSON), and the index options are a
single shared constant because loadJson requires the exact options the
index was serialized with.

Warm search path (load cached index + query) drops from ~36ms to ~14ms on
a 5,000-note vault. minisearch remains a devDependency as the migration
test oracle.

require("../package.json") cannot resolve inside a bun-compiled binary's
$bunfs, which broke every build:bun executable at startup. A static JSON
import bundles the version at compile time and behaves identically under
node and tsc output.

The keyword-extraction doc catches up with the pipeline it describes:
frontmatter-value weighting, the full structured-noise list with the
necessary-condition guards, the homogeneous-sibling collapse step, keyword
corroboration filtering, and a new caching section. configuration.md gains
the overview.collapse key and the auto-managed cache files. CLAUDE.md and
the README stop claiming .napkin/ holds config only, and the README states
the native engine's platform support.