Drawer search: loading feedback + relevance ordering everywhere

[sysread]

SYNOPSIS

Every drawer-side search now shows a Scanner spinner while
the embed+query round-trip is in flight and ranks results by
closest match. Chat, wiki, journal, recipes - one shape.

PURPOSE

Currently the four drawer searches disagree on feedback and
order:

• chat: spinner only on first-result-empty; stale results stay during subsequent searches
• wiki: spinner only on first load; an active query is forced back to alphabetical, throwing away the cosine ranking
• journal: substring only over in-memory entries; no spinner
• recipes: substring only over in-memory titles; no spinner; no embedding pipeline at all on the server side

Bad because the user can't tell working from frozen between
keystrokes, and the closest-meaning match doesn't surface.

DESCRIPTION

Three layers.

(1) How it works today.

• Chat search runs embed → searchThreads (exact-title hits then
  semantic by similarity); the Scanner is gated on searchResults.length === 0 so subsequent searches read as frozen.
• Wiki sidebar runs embed → searchWikiArticles (semantic merged
  with ILIKE) but re-sorts the result list alphabetically at the
  view layer, and the loading text only fires on first mount.
• Journal sidebar filters journal.entries in-memory by substring
  over content/mood/topics. No async, no spinner.
• Recipe sidebar filters cookbook.recipes in-memory by title
  substring. Schema has no embedding column on recipes (the
  comment in schema.sql:889 says "personal cookbook is small, ILIKE
  is fast enough").

(2) What this PR changes (same names, same order):

• Chat: Scanner now replaces any prior result list whenever
  searchBusy is true, not only when results are empty. Ordering
  untouched.
• Wiki: alphabetical sort runs only when query is empty; an
  active query passes the server's cosine order through. Scanner
  replaces the listing for every in-flight search, not just the
  initial load.
• Journal: rewritten to embed + searchJournalEntries (the same
  function the journal_search tool uses). Per-entry hits are
  aggregated by date keeping the best per-date similarity as
  the date's rank. Scanner during in-flight; empty-query browse
  mode unchanged.
• Recipes: added the full embedding pipeline on the server side
  (no prior infra existed). New columns on public.recipes
  (embedding vector(2048) + claim cols), clear_recipe_embedding_on_change
  trigger, three RPCs (claim_next_pending_recipe,
  save_recipe_embedding_if_claimed, search_recipes_by_embedding),
  one new worker source (src/lib/embeddings/sources/recipes.ts)
  registered in worker.ts, one new service method
  (SupabaseService.searchRecipes) that merges semantic+ILIKE the
  same way searchWikiArticles does. Sidebar consumes it. Sort
  picker hides during a search since relevance is the active
  sort then. recipe_list / recipe_search tools are
  untouched - still ILIKE-on-title; the new pipeline only feeds
  the human drawer search.

(3) How this fixes PURPOSE.

Every search surface now uses the same idiom: type → debounce →
spinner replaces list → ranked results arrive. The closest-meaning
match is at the top in all four tabs.

Notes:

• recipes ship graceful-degrades before the schema sync lands -
  searchRecipes falls back to ILIKE-only when the new RPC isn't
  present yet, matching the wiki/journal "no Venice client" path.
• the schema change is idempotent (every statement is add column if not exists / drop ... if exists / create or replace) so
  mise run sync is safe to re-run.
• the existing "No embedding column" comment block on recipes
  is updated rather than deleted - the rationale for the original
  decision is still relevant context (the LLM tool path still uses
  ILIKE; the new pipeline is for the human drawer).
• the journal sidebar's previous substring-on-three-fields
  matching is replaced with the server-side semantic+ILIKE merge,
  which is a behavioural change (a query that matched a topic
  substring but not content might rank differently now). The
  trade-off was accepted as part of choosing the "Add semantic
  search everywhere" option.
• docs: docs/user/search.md was a stub (H1 + placeholders); now
  fleshed out per-tab. docs/dev/cookbook.md and docs/dev/embeddings.md
  updated to mention the recipes source.

Test plan

• open each drawer tab, confirm Scanner replaces the list while
  typing and results appear in relevance order
• empty-query browse mode unchanged in all four tabs (wiki
  alphabetical; recipes by 'updated' or 'rating'; journal newest-day
  first; chat by Recent/Older/Archived buckets)
• mise run sync applies the recipe-embedding schema cleanly
• background embeddings worker picks up the recipes source on
  next sweep
• recipe-list LLM tool still ILIKE-on-title (unchanged)
• mise run check green

---

Generated by Claude Code