Support overriding max_tokens for OpenRouter models

[erendm-alt]

Current behavior

When running OpenWiki with OPENWIKI_PROVIDER=openrouter, the agent loop always requests the
model's advertised maximum output tokens (profile.maxOutputTokens from
@langchain/openrouter's model-capability table) as max_tokens on every request, regardless of
how much output is actually needed for the task.

For example, with OPENWIKI_MODEL_ID=z-ai/glm-5.2, every request sends "max_tokens": 65536,
because that's the model's ceiling in @langchain/openrouter's bundled profile — not because
OpenWiki, the workflow, or the caller asked for that much.

Why this is a problem

This prevents using OpenWiki on lower-credit OpenRouter accounts even when a much smaller output
budget (e.g. 8192 or 4096 tokens) would be entirely sufficient for the documentation update being
generated. OpenRouter rejects the request outright with a 402 ("This request requires more
credits, or fewer max_tokens...") rather than truncating or degrading gracefully, so the run fails
completely instead of just producing a smaller update.

We traced this end-to-end (workflow env vars → OpenWiki source → @langchain/openrouter's request
building) and confirmed OpenWiki itself never sets a maxTokens field when constructing the model
client (ChatOpenRouter in src/agent/index.ts), and there's no OPENWIKI_MAX_TOKENS-style env
var, CLI flag, or config key anywhere in the codebase to override it.

Desired behavior

Allow an explicit max_tokens override for OpenRouter (and ideally other providers) via one of:

• an environment variable (e.g. OPENWIKI_MAX_TOKENS)
• a config file setting
• a CLI flag

so accounts with a smaller credit balance can still run OpenWiki against models with a large
advertised output ceiling, by capping the per-request budget explicitly.

Environment

• openwiki 0.3.0
• @langchain/openrouter 0.4.5
• Model: z-ai/glm-5.2
• Provider: openrouter