Search & recommendations upgrade + soft-bundle upsells

Why / Context

MVP discovery is Postgres FTS (GIN indexes on Course/Project) — fine for launch scale, weak on typo tolerance, faceting speed, and ranking. Marketplace network effects (teardown §5.3: learners who came for one creator discovering others) need real search + cross-creator recommendations + the PRD §6.5 "soft bundle" upsells at checkout (never forced).

Schema already in place — NO migration required

• Course/Project catalog fields (tags, category, level, tier, ratingAvg, enrollmentCount) are the index document source
• Bundle/BundleItem + Price(BUNDLE) for soft-bundle upsells (tables dormant since MVP)
• OrderItem history powers "learners also bought" co-purchase signals

Implementation plan

1. Pick engine: Typesense Cloud vs Meilisearch (decision doc: typo tolerance, facets, geo-free, cost) — default Typesense.
2. Index pipeline (src/lib/search/indexer.ts): denormalized docs for courses/projects/creators; sync on publish/update/archive (after-write hooks in the publish actions) + nightly full rebuild function.
3. Search API + UI: instant search with facets (category, level, tier, price band, rating, language, creator), typo tolerance, synonyms (js/javascript); replace FTS route while keeping FTS as fallback flag.
4. Ranking: tune with weights (title > tags > description) + popularity boosts (enrollmentCount, ratingAvg with Bayesian floor).
5. Recommendations v1 (SQL, no ML): "learners also bought" from OrderItem co-occurrence; "more from this creator"; "next after this course" from category+level ladders — precomputed nightly into a cache table? No — keep in a materialized view via ADDITIVE migration if needed (allowed) or Redis-free JSON cache on the doc.
6. Soft-bundle upsells: creator defines Bundles in studio; checkout shows "add the mentor-guided capstone for this course at X% off" (PRD's exact pattern) — cart applies bundle discount when both members present.
7. Search analytics: log queries/zero-result terms (PostHog, issue #12) to feed synonyms and content gaps.
8. Storefront-scoped search (tenant filter) reusing the same index.

Acceptance criteria

• Typo'd query ("javascrpt") returns relevant results with facets under 100ms p95 (engine-side)
• Publishing/unpublishing reflects in search within a minute
• Course detail shows "also bought" and creator cross-sell modules
• Checkout surfaces a soft-bundle suggestion when applicable; discount applies correctly and remains optional
• Zero-result queries logged and reviewable

Dependencies / sequencing

• MVP M6 (catalog) required; issue #12 for analytics loop
• Bundles UI unlocks the dormant Bundle tables

Start by reviewing the implementation plan, the existing catalog data, and src/lib/search/indexer.ts, then trace the publish actions and dependencies on MVP M6 and issue #12. Done means the listed search, recommendation, optional bundle-discount, and zero-result analytics acceptance criteria are met without a migration unless the plan requires one.