fix(ai): classify Novita context length rejections as context overflow - #52134

Problem

Novita rejects an oversized prompt with HTTP 400 and no token counts:

{"message":"The input is longer than the model's context length trace_id: 39d8c3a5c6c91dbc8dd9055f0b37e084","type":"invalid_request_error"}

The existing pattern for this wording required token counts (The input (N tokens) is longer than the model's context length (N tokens)), so it didn't match. It was classified as a plain InvalidRequest, compaction never started, and long sessions failed.

Fix

Make the token counts optional in that existing context-overflow pattern in packages/ai/src/provider-error.ts. The wording with counts still matches.

Verification

• Live, a 1.3M-token prompt to deepseek/deepseek-v4.1-flash and meta-llama/llama-3.3-70b-instruct now returns 400 InvalidRequest: context-overflow, not retried.
• New test uses the captured body. The existing test for the counted wording still passes.
• bun test in packages/ai and bun run check pass.

Not fixed here

qwen/qwen3.8-flash rejects the same prompt with only invalid request error trace_id: ..., which has nothing to classify on.