#2629 GCP codegen: filtered generate:gcp <service> produces different output than a full generate:gcp run

Found while working on swamp-club #2287.

With the same fetched discovery schemas, running deno task generate:gcp compute changes 118 compute models, while a full deno task generate:gcp immediately afterwards reverts most of them. The filtered run picks up a different compute API schema: for example compute routes and backendservices gain a returnPartialSuccess query parameter in LIST_CONFIG, plus a CalVer bump and a No schema changes upgrade entry. The full run does not include it. Running the full generate twice is idempotent, so the difference is between the filtered and unfiltered code paths.

Likely cause: fetch-schema:gcp also fetched additional API versions (Fetched 52 additional API versions), and the filtered path appears to select a different version document for the service than the full path.

Impact: the documented workflow (filtered generate first, then full) leaves spurious diffs and version bumps if someone commits after the filtered run, and the nightly full regeneration would then revert them. Filtered and full generation should select the same schema version per service.