Classify Anthropic's "input length and max_tokens exceed context limit" as ContextLengthExceededError - #907

What this does

Adds /context limit/i to ErrorMiddleware::CONTEXT_LENGTH_PATTERNS so Anthropic's second context-window error raises RubyLLM::ContextLengthExceededError instead of BadRequestError:

{"type":"error","error":{"type":"invalid_request_error","message":"input length and max_tokens exceed context limit: 900000 + 128000 > 1000000"}}

Closes #906.

Why

Anthropic checks input + max_tokens <= context_window before it checks the prompt alone, and the chat payload always sends max_tokens (model.max_tokens || 4096, 128k for claude-opus-4-8). So on a long-context model this is the first error a growing conversation hits, at roughly context_window - max_output_tokens input tokens. The "prompt is too long" pattern from #755 only fires later, once the prompt alone no longer fits. Applications rescuing ContextLengthExceededError for Anthropic therefore missed the case that actually occurs.

Type of change

• Bug fix

Scope check

• I read the Contributing Guide
• This aligns with RubyLLM's focus on LLM communication
• This isn't application-specific logic that belongs in user code
• This benefits most users, not just my specific use case

Quality check

• I ran bundle exec rspec spec/ruby_llm/error_middleware_spec.rb spec/ruby_llm/error_spec.rb (48 examples, 0 failures) and bundle exec rubocop on the touched files
• I added a spec for the new phrasing next to the existing "prompt is too long" one
• overcommit --run — not run locally (mysql2 does not build on this machine); RuboCop and the affected specs were run by hand

AI-generated code

• I used AI tools to help write this code
• I have reviewed and understand all generated code (required if above is checked)

API changes

• Breaking change
• New public methods/classes
• Changed method signatures

Related issues

#906 (this), #755 (the earlier "prompt is too long" half). The streaming-status half of #906, where 1.16.0's Anthropic adapter labels every failure a 500, is already fixed on main by failed_http_status; this PR only adds the missing pattern.