Cut the faceted page to a single engine round-trip (overlap + ride the listing search)

[ddeboer]

Context

After #532 (PR #554), a faceted page costs 2 engine round-trips: the listing search, then — only once it resolves — the batched facet multi_search. The facet queries derive solely from the resolved SearchQuery, not from the listing result, so the second round-trip need not wait for (or even be separate from) the first.

Proposal

Two composable steps, in order of effort:

1. Overlap: let the root resolver return its source synchronously with the listing search as a pending promise (items/total sub-resolvers await it), so the facet dispatch fires in the same tick and the two round-trips run concurrently — max() instead of sum.
2. Merge: facets whose effective where (after skip-own-filter) equals the listing query’s where can ride the listing search via facet_by — the common unfiltered browse becomes 1 round-trip. Needs the root resolver to read the facet selection from GraphQLResolveInfo (including fragments), which ADR 4 anticipates (‘selection is the request’).

Acceptance

• Unfiltered browse: 1 Typesense round-trip (listing + facets in one search); filtered pages: concurrent, not sequential, round-trips.
• Facet counts, skip-own-filter semantics, per-facet failure isolation, and the GraphQL contract unchanged.

Notes

• Deferred from Batch facet searches into a single multi_search (per-facet engine.search fan-out) #532; see the ‘Deliberately not done’ notes in ADR 5.