lsm / neokai / 28214505273 / 26

82%

dev: 82%

push

github

fix: classify GLM 1305 overload as retryable via isRetryableProviderError (#2183)

GLM (Zhipu) 529 `[1305][该模型当前访问量过大，请您稍后再试]` ("model traffic
too high, try again later") was still terminal: the HTTP 529 transport status
often does not survive into the surfaced error string, and dev's
RETRYABLE_PROVIDER_ERROR_TEXT had no GLM-specific signals, so
isRetryableProviderError() returned false and the bounded retry path (#2184)
never fired.

Add the GLM overload signals to RETRYABLE_PROVIDER_ERROR_TEXT:
- `[1305]` — bracketed payload shape only, so unrelated numerics (token
  counts, ports, request ids) cannot false-positive into a retry.
- `访问量过大` / `当前访问量过大` — overload-specific Chinese descriptions
  ("access volume too high") that catch the payload regardless of code
  rendering.

Generic advice like 稍后再试 ("try again later") is intentionally excluded —
it appears on localized terminal/validation errors too. The bracketed-code
guard mirrors dev's word-boundary numeric philosophy.

Rebased onto dev: the generic 5xx / bridge-5xx / overload / case-insensitive
handling landed in #2184, so this is now the focused non-redundant delta.

11382 of 16404 branches covered (69.39%)

Branch coverage included in aggregate %.

14216 of 17932 relevant lines covered (79.28%)

54.26 hits per line