RG manifest: widen eligibility to vocab-term covers (regions, bboxes)

`include` lists on rides-v5 are always vocab terms (`s:` passthrough or
`vocabCover` translation), and vocab-pyramid avail covers are exact-match
terms too — so the per-token manifest predicates are valid for every
shape, not just `s:`-key covers. Measured region covers are ≤16 tokens
(NYC = 16 vocab cells, 0 station keys), inside the 45-token predicate
cap; dim-filtered rides queries and avail excludes keep the legacy
guarded path. Verified byte-equal to the footer path on all three region
± covers (rides) and a mixed cells+`s:` cover (avail-v6), 1.4-3.2×
faster cold.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: ryan-williams

gbfs/api/src/avail_geo.ts

@@ -600,9 +600,18 @@ async function serveGeoReduced(
 	// back to the footer path per key and lazily fill (big shards only —
 	// see `MIN_FILL_RGS`; the Lambda-churned tip shards stay on the cheap
 	// fallback so their constant rewrites don't grind D1 writes).
+	// Manifest fast path (`specs/rg-manifest.md` P3): explicit covers on
+	// vocab pyramids use exact-match per-token predicates — valid for any
+	// vocab term (`s:` keys AND vocab cells; rows exist only at vocab
+	// terms, so a non-vocab token matches nothing on either path).
+	// Region-scale vocab covers measure ≤16 tokens (NYC), well inside the
+	// predicate builder's 45-token cap. Excludes stay on the legacy path
+	// (query-time subtraction; rare). Non-vocab pyramids (avail-v3) keep
+	// the legacy path: their S2 cover cells match rows by containment,
+	// not equality, so per-token exact predicates would undercount.
 	const manifestEligible = userCells !== null
 		&& userCellsExclude.length === 0
-		&& userCells.every((c) => c.startsWith('s:'));
+		&& (PYRAMIDS[pyramidName]?.vocab === true || userCells.every((c) => c.startsWith('s:')));
 	// Thread `binCol` + per-segment range so hyparquet prunes row groups by
 	// `dt` column stats — shards are `(dt, s2_cell)`-sorted with small
 	// row groups, so a sub-shard time window reads only the matching RGs.

gbfs/api/src/rides_v1.ts

@@ -741,14 +741,17 @@ export async function serveRidesV5(
 			registered,
 		);
 	const rgFilters = parseDimFilters(url) ?? [];
-	// RG-manifest path (`specs/rg-manifest.md` P1): `s:`-key covers with no
-	// dim filters serve from the D1 row-group index — no footer parse, so
-	// segments fan out in parallel and the footer guard doesn't apply
-	// (misses fall back to the guarded footer path per key, and fill the
-	// manifest via `defer`). Dim-filtered or vocab-cell queries keep the
-	// legacy guarded sequential path (dim RG-prune semantics + large
-	// token sets are out of P1 scope).
-	const manifestEligible = rgFilters.length === 0 && include.every((c) => c.startsWith('s:'));
+	// RG-manifest path (`specs/rg-manifest.md` P1+P3): serve from the D1
+	// row-group index — no footer parse, so segments fan out in parallel
+	// and the footer guard doesn't apply (misses fall back to the guarded
+	// footer path per key, and fill the manifest via `defer`). `include`
+	// is always a vocab term list here (`s:` passthrough, or raw
+	// covers/bboxes translated positive-only via `vocabCover`), so
+	// per-token exact-match predicates are valid for every shape;
+	// region-scale covers measure ≤16 tokens (NYC), inside the predicate
+	// builder's 45-token cap. Dim-filtered queries keep the legacy
+	// guarded sequential path (dim RG-prune semantics are out of scope).
+	const manifestEligible = rgFilters.length === 0;
 	let shardRows: Row[][];
 	try {
 		if (manifestEligible) {