fix(search): self-describing pagination cursors so load-more can't diverge from first-load

[tompscanlan]

Problem

loadMoreEvents picked its pagination backend implicitly: if a search term was set and a Meili backend was configured it paginated via Meilisearch (cursor = a numeric offset), otherwise via D1 listRecords (cursor = a base64url keyset).

That breaks whenever a page's first load comes from one backend but its load-more resolves to the other — the two cursor formats are incompatible:

• D1 keyset → Meili: Number(base64url) → NaN → offset 0 → a relevance-reordered duplicate of page 1.
• Meili offset → D1 listRecords: the offset is ignored and the discoverable / time-bound filters the first page applied get silently dropped, so later pages drift into past / non-discoverable events.

The topics page tripped exactly this; the search page only avoided it by being end-to-end consistent (Meili+Meili, or cursor:null in its D1 fallback).

Fix — self-describing cursors

Tag every client-facing cursor with the backend that issued it, and route load-more by that tag instead of re-deriving the backend from request shape.

• New lib/contrail/cursor.ts — tagCursor(backend, raw) / parseCursor(cursor). Prefix scheme meili: / d1:; the : separator is collision-free because a base64url keyset and a decimal offset both exclude :. A null/empty raw cursor stays null (never manufactures a cursor). Untagged input → { backend: null, raw } for the legacy fallback.
• events-load-more.ts — routes on the cursor's tag. d1-tagged → D1 pipeline even under search + configured Meili; meili-tagged → Meili; untagged → the old inference (for cursors in flight across the deploy). A meili-tagged cursor with no usable search context fails safe to { events: [], cursor: null } rather than restarting page 1 on the wrong backend. D1 output cursors are d1-tagged; the inbound tag is stripped before the D1 read.
• search/server/query.ts — hydrateToPage emits a meili:-tagged offset; parseOffsetCursor accepts a tagged or bare offset and maps a d1-tagged (or NaN) cursor to offset 0 (clean restart, never NaN-from-base64url).
• First-page routes (events, p/[actor]/hosting, p/[actor]/past-events) — emit d1-tagged first-page cursors and untag any inbound deep-link cursor before the D1 read.
• search + topics — keep cursor:null (their fetchParams carry no pipeline, so a resumable D1 cursor would still drop the discoverable + startsAtMin filters). Deep pagination for these is deferred to a separate task. Stale "re-routes to Meili whenever a backend is configured" comments were rewritten to describe tag routing.

Verification

• pnpm test (apps/web, vitest): 183 passed / 4 skipped (187 total), 21 files — green.
• pnpm run check (svelte-check): 0 errors (7 pre-existing style warnings in unrelated files).
• New unit tests cover: tag round-trip for both cursor kinds; both divergence directions (D1 first page + search + Meili → stays D1 with filters intact; Meili first page → stays Meili); the two fail-safe cases; and the legacy untagged-cursor path.

Branch is 4 commits ahead of main, 0 behind; merges cleanly.