[felipebalbi]

Summary

Moves job fuzzy search — plus the "↑ N new" pill and the live poller's
change-detection — off the full-history in-memory JobIndex in paavo-web
and into SQLite, so the web process no longer holds the whole job history in
RAM just to search it. SQLite drives iteration and returns only the requested
page.

• Ranking + membership in SQL. A custom fuzzy_score SQLite scalar function
  (SkimMatcherV2, registered on every connection) ranks matches; a subsequence
  LIKE '%a%l%m%c%x%' pre-filter decides membership. So almcx still matches
  and ranks alice … mcxa266, identically to the old in-memory matcher — but the
  expensive UDF runs only on matched rows, and the pagination COUNT(*) calls it
  zero times.
• JobIndex deleted. /api/jobs reads SQLite directly (fuzzy page +
  search_count, or an as_of-pinned time-ordered page for a blank query); the
  "N new" pill is a COUNT(*) WHERE submitted_at > as_of; the poller's
  change-detection switches to a bounded GROUP BY state activity_digest
  (no full-table materialization). fuzzy-matcher moves paavo-web → paavo-db.
• SPA debounce 150 → 250 ms, since each committed keystroke now triggers a
  DB-side scan.

No schema migration, no paavo-proto / wire change, no paavod change, no
/api/events contract change. Design + plan: docs/superpowers/specs/2026-06-17-db-side-fuzzy-search-design.md,
docs/superpowers/plans/2026-06-17-db-side-fuzzy-search.md.

Notable design points / trade-offs

• Count ≡ page is structural: search_count and search_index_page use a
  byte-identical lowercased haystack and the same subsequence pattern, so a page
  can never reference a row the count excludes. A divergence-guard test asserts
  the LIKE membership set equals the fuzzy_score IS NOT NULL set.
• Accepted (documented) trade-offs: search is an honest O(n) table scan per
  committed (debounced) keystroke — fine at lab scale; the activity_digest can
  miss two offsetting same-tick state transitions and self-heals on the next poll
  (it's a refetch hint, never a correctness input).
• Cross-branch note: if feat/dashboard-sql-aggregates lands after this, its
  "Recent activity" list should read JobRow::list_index_page(None, 0, 8) rather
  than reintroduce JobIndex.

Test plan

• cargo fmt --all -- --check — clean
• cargo clippy --workspace --all-targets -- -D warnings — clean
• cargo test --workspace — green (362 passed, 0 failed; 6 hardware/tooling-gated ignores)
• New coverage: db-level fuzzy match/rank, list ordering + as_of pin,
  count_newer boundary, activity_digest change-detection, the
  LIKE≡fuzzy_score divergence guard, search pagination; /api/jobs
  end-to-end through SQL (api_jobs); digest-driven SSE bump (api_events).
• trunk build --release (crates/paavo-web-ui) — builds (workspace-excluded wasm32)
• Manual smoke (human): PAAVO_FAKE_RUNNER=1 daemon + paavo-web; on
  /jobs, typing almcx ranks an alice … mcxa266 row first, clearing the
  box returns to the time-ordered list, the "↑ N new" pill updates on new
  submissions, and the query fires only ~250 ms after typing stops.