KEY FIX: GLM returns drained-key errors (code 1113) with HTTP 429 status,
not real rate-limits. The proxy was short-banning (20s) drained keys so
every request re-walked ~640 dead keys. Now body is inspected: 1113 ->
1h ban, genuine 429 -> 20s. glm-4.5-air now 0.5s, glm-4.6 ~3-5s (its
own reasoning floor).
Shared connection pool, bootstrap classify-all on startup + recovery-only
sweeps, tiered ban TTLs, multi-harness (OpenAI+Anthropic+Gemini paths),
round-robin + per-key retry, encrypted models.json (AES_KEY), instant
/health, / and /v1/models public, docs blocked.
