Write pipeline for POST / PATCH / PUT / DELETE on a relation.

Mirrors PostgREST's mutation semantics:

• POST — INSERT ... RETURNING, optionally ON CONFLICT for upsert (Prefer: resolution=merge-duplicates|ignore-duplicates, ?on_conflict=).
• PATCH — UPDATE ... WHERE <filters> RETURNING, honoring the request's column filters.
• DELETE — DELETE ... WHERE <filters> RETURNING.
• PUT — single-row upsert keyed by the request's PK filter (INSERT ... ON CONFLICT (<pk>) DO UPDATE).

Row values arrive as a JSON payload and are expanded into typed columns through Postgres' jsonb_array_elements (one bound jsonb parameter), so a string like "{1,2,3}" coerces to integer[] exactly as PostgREST does; user data is never interpolated. The ?columns= param selects which payload keys become target columns (others are ignored); Prefer: missing=default fills the column DEFAULT for keys a row omits.

The Prefer: return= token selects the response shape; Preference-Applied echoes the honored tokens in PostgREST's canonical order.