fix: a non-JSON vendor error body keeps the vendor's status; cut dead items and wrong comments

[CMGS]

Four commits from a whole-repo read of main@38927996: two fixes, then the read's cut-list.

Fixes

a3e4d37c — a non-JSON vendor error body keeps the vendor's status. An upstream that answers an error status with a non-JSON body (a CDN or proxy HTML page on 502/504, a relay's plain-text 413/429) failed serde parsing and became a gateway-internal 500 parse … response. The account took a fault and failed over, the model's fallback_models chain never ran (is_upstream_fault ignores a 500 without an upstream status), a client-caused 413 page cooled the account, and the client saw internal_server_exception instead of the documented 424 model_error_exception with original_status_code (docs/api.md, docs/governance.md "Model fallback"). The six body-parse sites now classify such a body through vendor_error at its real status; a non-JSON body on a success status keeps the gateway parse error. Hot path unchanged: only the error branch differs.

26dcb85e — the realtime refusal on the chat path names the model. Since 39c81a2 the 400 interpolated the protocol string, so it always read "realtime model `realtime`".

Cuts (behavior-preserving, no test changed)

• ca48cfe2 — BillingRecord derives only Serialize, so its seven #[serde(default)] never took effect; the mock's meta.llama arm is unreachable behind the /model/ arm.
• 740f67fe — comments the code contradicts or repeats: MOCK_B64 is base64 of "MOCKBYTES", not PNG bytes; Account holds seven strings and two vectors; the content_block_delta arm stated one fact twice.

Net: prod −3 lines, comment lines −3, tests +38.

Evidence

• Failing-first tests: a_non_json_error_body_keeps_the_vendor_status (handler/tests/fallback.rs) failed on main with GatewayError { code: 1000, http_status: 500, message: "parse openai response", original_status: None }; a_realtime_refusal_names_the_model failed with "realtime model `realtime` …".
• Per commit on macOS: fmt and clippy clean; tests 687/0/4, 688/0/4, 688/0/4, 688/0/4 (passed/failed/ignored).
• Linux arm64, rustc 1.98.0, with Postgres 16 and Redis 7: fmt and clippy clean, 688/0/4; the Redis/Postgres suites ran (a gwtest_* database and governance keys were created).
• Live before/after with real vendor keys. A local fake proxy answered an HTML 502 page and a text 413; the fallback targets were real OpenAI and Anthropic models:

case	main	this branch
chat, proxy 502 page, fallback gpt-4o-mini	500 internal_server_exception	200 served by gpt-4o-mini, requested name echoed
/v1/messages, proxy 502 page, fallback claude-haiku-4-5	500	200
/v1/responses stream, proxy 502 page, fallback gpt-4.1-mini	error frame internal_server_exception	native stream from gpt-4.1-mini
3× chat, proxy 413 text, fallback configured	500 ×3, account cooling	424 model_error_exception, original_status_code: 413, no fallback, account ok
ledger	0 rows	3 rows, requested → served

• Full live matrix on the branch: 209/218. All nine failures are vendor-side: Ernie account overdue (2), SiliconFlow balance exhausted (402, 4) plus its rerank 503 once the 402s cooled that account, Brave key invalid (422), and one qwen-plus implicit-cache miss (an A/B/B/A rerun hit 4/4 on both binaries). The xAI and Ollama groups did not run (no key, no local instance).