Custom endpoints behind a gateway never get a context length, so Claude assumes 200k #754

On Codeman 1.47.0, a custom model endpoint that is a gateway in front of llama-swap (for example Bifrost) has no context length saved, so CLAUDE_CODE_MAX_CONTEXT_TOKENS is never set. Claude Code then prints Model "..." isn't recognized, context assumed to be 200k tokens and does not compact before a smaller server overflows. Codex, OpenCode and DeepSeek get no limit either.

Cause: modelContextLengths is only filled from llama-swap's /running launch command (and /props), which exists only for a loaded model and only when Codeman talks to llama-swap directly. A gateway has no /running. Discovery never reads the size from GET /v1/models, although current llama-swap sends context_length and context_window there (see llama-swap api.go) and Bifrost passes context_window through (Bifrost openai/models.go). OpenRouter-style servers send context_length.

Expected:

• Discovery reads a positive integer context_length, else context_window, from each /v1/models entry. The llama-swap /running value keeps priority for a direct llama-swap endpoint.
• An optional per-endpoint defaultContextLength is used for models whose size cannot be discovered, so a CLI never falls back to its own large default. Discovery must never overwrite it.
• Regression tests for both deployments (direct llama-swap and a gateway).