feat(marquee): Phase 2 joins + aggregation via the named-operation port

[2bTwist]

Stacked on #51 (Phase 1). Phase 2 is the first "SDK gap" phase, and the finding is that there is no core gap: joins and aggregation ride the existing named-operation port (store.run), whose calling convention is portable but whose implementation is per-backend. That is exactly the seam the core reserves for genuine divergence. The capability flags declare it; the UI surfaces it.

Schema (both backends)

people + credits make cast/director a real many-to-many relation. Supabase adds FK constraints (so PostgREST can embed) and a genre_counts GROUP BY view.

Named queries: same shape, divergent implementation

• movieCredits (the JOIN): Supabase = a server-side PostgREST embed over the FKs; Convex (convex/marquee.ts) + memory = follow refs by hand.
• genreCounts (the AGGREGATION): Supabase = a GROUP BY view; Convex + memory = scan and tally (the cost the aggregations: false flag warns about, exercised for real at Phase 5 scale).

makeBackend wires the typed MarqueeSchema with per-backend query impls (src/lib/enrich.ts, env-free so tests build typed backends directly). The catalog functions became generic over the schema so the bare test backends still fit.

UI (the thesis, made visible)

MovieDetail shows the real joined cast (director + billed actors with characters); the catalog shows per-genre counts. Both render an understated caption declaring HOW the data loaded on the active backend: "Cast joined server-side" vs "assembled by following references"; "Counted with a SQL view" vs "by scanning".

Verification

• test/enrich.live.test.ts: 12 tests (join order + name resolution, empty cast, aggregation, capability declaration) green on memory + Supabase + Convex. Catalog gate still 20/20. Full suite 32/32.
• Playwright UI smoke confirmed the cast join and the genre-stats strip with the correct divergence captions on both live backends, 0 console errors.

No packages/* change (the named-op port already covered it), so no changeset. avg-rating aggregation defers to Phase 3 (reviews need auth); Phase 2's aggregation is genre counts over existing data.