cloud-api: modelRateLimits publishes a 1M window for qwen3.8-flash while the proxy enforces 262144 #53

Found while fixing the published client configs (#8, #9), and raised by the #52 blocking reviewer.

The disagreement

internal/handlers/usage_quota.go (modelRateLimits):

// qwen3.8-flash: external model served via DashScope. YaRN extends the
// native window to 1M, which is what we expose to members.
"qwen3.8-flash": {ContextWindow: 1_048_576, Concurrency: 5, FillsPerMin: 6},

The community LiteLLM deployment:

model_name     | max_input_tokens
qwen3.8-flash  | 262144

And the model card on /docs/models: "262K token context, the model native window", with
{ label: Context, value: 262K tokens }.

So two of three sources say 262K and the rate table says 1M.

Why it is not cosmetic

modelRateLimits.ContextWindow is not only descriptive — the file says the effective ITPM is
ContextWindow * FillsPerMin. At 1_048_576 x 6 the model is granted ~6.3M tokens/min; at its
enforced 262_144 x 6 it would be ~1.6M. The member cannot actually spend the difference,
because the proxy rejects anything over 262_144 input, so the generosity is unreachable — but
any future reasoning that treats that ITPM as real (a comparison across models, a capacity
estimate) starts from a 4x overstatement.

It also makes the comment above it false: we do NOT expose the YaRN-extended window; the
deployment caps at native.

What to decide

1. If the intent is to serve 1M, the deployment needs max_input_tokens raised (and the model
   card updated) — DashScope has to actually accept it.
2. If the intent is native 262K, modelRateLimits should say 262_144 and the comment should go.

Option 2 matches what is served today.

Not changed here

/docs/examples now publishes 262144 for this model, following the proxy, and the test that
pins it carries this issue number as the reason it disagrees with modelRateLimits. Also
noted: mimo-v2.5 is 1_050_000 there against a 1M model card and no enforced cap — same
direction, 1424 tokens, not worth a change on its own.