fix: clamp max_tokens to the context window for OpenAI-compatible providers

[Sayt-0]

Summary

A self-hosted vLLM user (#3387) got "the conversation has exceeded the model's context window" on a bare "hello". Root cause: max_tokens (the per-response output budget) was set equal to the context window and forwarded verbatim. vLLM requires prompt_tokens + max_tokens <= context_window, so no room was left for the prompt.

Root cause

Cause	Handled by	Status
max_tokens forwarded unclamped	clamp in OpenAI client	fixed
YAML context_size read as 0 (uint64 not handled)	shared parser handles unsigned ints	fixed
Misleading config, no early signal	load-time warning + schema doc	fixed

Changes

• pkg/model/provider/openai/client.go: clamp max_tokens to window - 1024 when the window is known (provider_opts.context_size first, then models.dev), on both the chat-completions and responses paths. Window unknown, value left unchanged.
• pkg/config/latest/types.go: ContextSizeFromProviderOpts, a single parser now used by the runtime and the client. Handles uint64/uint/uint32 (goccy/go-yaml decodes positive YAML integers as uint64), so a YAML context_size is honored. This also restored proactive compaction for those configs.
• pkg/runtime/session_compaction.go: providerContextLimit delegates to the shared parser.
• pkg/config/max_tokens_warning.go: load-time warning when max_tokens >= context_size, or when it is set to a context-window-sized value with no discoverable window.
• agent-schema.json: max_tokens documented as the output budget, not the context window.

Reproduction (mock enforcing vLLM's rule, max_model_len = 262144)

Scenario	max_tokens sent	Server	Result
Reporter config (no context_size)	262144	reject 12 + 262144 > 262144	bug reproduced, warning shown
With context_size: 262144	261120 (clamped)	accept 12 + 261120 <= 262144	reply returned

Design note (open to maintainer preference)

The clamp reserves a fixed 1024-token headroom (window - 1024), matching the Anthropic client's clampMaxTokens. It is deliberately prompt-agnostic (no token-count round-trip): it guarantees room for a small prompt, not necessarily a very large one. A large agent prompt under a known window could still overflow and would fall through to the existing overflow detection and compaction. If a percentage-based margin (for example window - max(1024, window/8)) is preferred to better fit large agent prompts, it is a one-line change.

Testing

• pkg/model/provider/openai: clamp fires when the window is known, verbatim when unknown, plus a unit test of the clamp math.
• pkg/config/latest: parser covers int/uint64/float64/string and a real YAML round-trip.
• pkg/config: load-time warning cases.
• pkg/runtime: uint64 case added to the existing context-limit test.
• task build and task lint pass.

Fixes #3387