OpenAPI spec documents endpoints that 404: /v1/model/{model_name}; two parallel list endpoints

Summary

The publicly-served OpenAPI spec at https://cloud-api.near.ai/api-docs/openapi.json documents endpoints that don't actually exist. A client generated from the spec hits bare 404.

Spec lies / unimplemented endpoints

GET /v1/model/{model_name} — documented but returns 404 size=0

curl -s -w "HTTP %{http_code} size=%{size_download}\n" \
  -H "Authorization: Bearer $KEY" \
  "https://cloud-api.near.ai/v1/model/openai/gpt-oss-120b"
# HTTP 404 size=0

Even though the spec lists this path with a documented schema (in paths.[/v1/model/{model_name}].get), the actual gateway returns a bare 404 — same family as #614.

Inconsistencies between spec and runtime

Two parallel model list endpoints

• GET /v1/models — works, OpenAI-compat shape ({object: "list", data: [...]})
• GET /v1/model/list — works, custom shape ({models, limit, offset, total})

Both are documented in the spec. The custom endpoint supports pagination; the OpenAI-compat one doesn't (silently ignores ?limit=). Two parallel endpoints with different shapes is confusing — pick one.

Inconsistent /v1 prefix

The spec documents:

• POST /conversations/{conversation_id}/archive (no /v1 prefix)
• GET /v1/conversations/{conversation_id} (has /v1 prefix)

Five /conversations/... paths are documented without the /v1 prefix; the rest have it. Probably a routing bug that landed in spec generation. (Didn't actually hit these — couldn't auth as a real user.)

Health version vs spec version disagree

/v1/health reports "version":"0.0.0" (already filed in #647), but the OpenAPI spec at /api-docs/openapi.json declares info.version: "1.0.0". Both should be the same source of truth.

Same raw-422 envelope-bypass as #609

POST /v1/score with missing model field returns:

HTTP 422
Failed to deserialize the JSON body into the target type: missing field `model` at line 1 column 13

Plain text, not the standard JSON envelope. Same pattern as /v1/rerank in #609 and frequency_penalty="oops" noted in #622.

Suggested fix

1. Either implement GET /v1/model/{model_name} or remove it from the spec.
2. Pick one list endpoint surface (/v1/models for OpenAI-compat OR /v1/model/list for custom-paginated) and deprecate the other; don't ship both.
3. Add the /v1/ prefix to the /conversations/... paths so the prefix is consistent.
4. Source /v1/health.version from the same value as info.version in the OpenAPI spec.
5. Apply the JSON envelope error handler to POST /v1/score too (same as #609 fix).

/api-docs/openapi.json itself works correctly (200, 206 KB, OpenAPI 3.1) — a good resource if it can be made source-of-truth.

---

Found during routine cloud-api smoke test.