fix(typegen): only collapse ROWS 1 return types when SetofOptions is emitted

[mandarini]

Fixes supabase/supabase#46525.

The TypeScript typegen drops the [] from the return type of every set-returning function declared with ROWS 1, but the compensating SetofOptions metadata is only emitted when the function returns a named table, view, or composite type. For ad-hoc RETURNS TABLE (...) and scalar SETOF functions this produced a single-object type while PostgREST still returns an array at runtime, and there is no metadata for postgrest-js to correct it. The collapse is now gated on the same condition as the SetofOptions emission, so those functions keep the array type, while named-relation functions (where postgrest-js restores the array for direct RPC calls via isOneToOne) are unchanged.

Note for consumers: regenerating types flips Returns from object to array for affected functions. The resulting type errors point at real runtime mismatches, since PostgREST has always returned an array for these calls.

[mandarini]

E2E verification against postgrest-js

I ran the cross-repo check locally: built two pg-meta images — one from this PR's merge-base (e7d86a3) and one from this branch — pinned each into the postgrest-js test stack, regenerated types for the postgrest-js test database with both, and compared against PostgREST's actual runtime responses.

1. No change to postgrest-js's existing type surface

supabase gen types typescript --local --schema public,personal output is byte-identical between the merge-base image and this PR's image, for the entire postgrest-js test database.

This is expected: all three ROWS 1 functions in that schema (get_user_profile_non_nullable, get_user_first_message, function_using_setof_rows_one) return named relations, so they keep the collapsed object Returns + SetofOptions.isOneToOne: true. The contract postgrest-js relies on (isOneToOne restoring the array shape for direct RPC calls) is untouched — no inference regression is possible against the current test surface.

2. For the broken shapes, the new types match runtime; the old ones don't

Added the affected shapes to the live test DB and called them through PostgREST:

Function	merge-base typegen	this PR	actual PostgREST response
RETURNS TABLE(id int, name text) ROWS 1	{ id, name } ❌	{ id, name }[] ✅	[{"id":1,"name":"hello_one"}]
RETURNS SETOF text ROWS 1	string ❌	string[] ✅	["hello_scalar"]

Neither shape emits SetofOptions, so postgrest-js has no metadata to correct the collapsed type — the merge-base types diverge from the actual response, which is exactly supabase/supabase#46525.

3. postgrest-js type suite (tstyche)

• Committed types.generated.ts (control): 29/29 files pass
• Fully regenerated types (either image — outputs are identical): 28/29, the one failure being the scalar_computed_count / scalar_computed_ids assertions in embeded_functions_join.test.ts

That failure is pre-existing drift, not this PR: it reproduces identically with the unmodified merge-base image. The committed types.generated.ts in postgrest-js contains SetofOptions.isNotNullable entries (introduced via supabase/supabase-js#2224) that aren't emitted by any pg-meta master build, so regenerating with any current master-based image drops them. Worth tracking separately from this fix.

Reproduction steps

# 1. Build both images so the Supabase CLI resolves them locally (checked before pulling)
git -C postgres-meta worktree add /tmp/pgmeta-base $(git -C postgres-meta merge-base HEAD master)
docker build postgres-meta -t public.ecr.aws/supabase/postgres-meta:rows1-pr
docker build /tmp/pgmeta-base -t public.ecr.aws/supabase/postgres-meta:rows1-base

# 2. In supabase-js/packages/core/postgrest-js: pin the image tag and start the stack
printf 'rows1-base' > test/supabase/.temp/pgmeta-version
pnpm exec supabase --workdir test start

# 3. Generate with each image (the pin is re-read per invocation)
pnpm exec supabase --workdir test gen types typescript --local --schema public,personal > /tmp/types-base.ts
printf 'rows1-pr' > test/supabase/.temp/pgmeta-version
pnpm exec supabase --workdir test gen types typescript --local --schema public,personal > /tmp/types-pr.ts
diff /tmp/types-base.ts /tmp/types-pr.ts   # → identical

# 4. Runtime source of truth for the affected shapes
docker exec supabase_db_test psql -U postgres -c "
CREATE FUNCTION public.e2e_rows1_ad_hoc_table(prefix text)
RETURNS TABLE(id int, name text) LANGUAGE SQL STABLE ROWS 1
AS \$\$ SELECT 1, prefix || '_one' \$\$;
CREATE FUNCTION public.e2e_rows1_scalar_setof(prefix text)
RETURNS SETOF text LANGUAGE SQL STABLE ROWS 1
AS \$\$ SELECT prefix || '_scalar' \$\$;
NOTIFY pgrst, 'reload schema';"
curl -s http://127.0.0.1:54321/rest/v1/rpc/e2e_rows1_ad_hoc_table \
  -X POST -H "apikey: $ANON" -H "Content-Type: application/json" -d '{"prefix":"hello"}'
# → [{"id":1,"name":"hello_one"}]

# 5. Type suite: swap regenerated types in, post-process, run tstyche
cp /tmp/types-pr.ts test/types.generated.ts
node scripts/update-json-type.js && pnpm exec prettier --write test/types.generated.ts
pnpm exec tstyche

Follow-up regression tests for the ROWS 1 array contract on the SDK side are tracked in SDK-1383.

[avallete]

Nice. I think we'll want the new tests to be added to postgrest-js against actual e2e postgrest runtime when this is merged and released. That way we don't get a regression on the behavior in the future 👍