fix: distinguish vLLM 'at least' lower-bound from exact input token counts

[jamesreader]

Problem

vLLM reports context window errors in two forms:

1. Lower bound: your prompt contains at least 81921 input tokens — this is an early tokenizer cutoff, NOT the actual prompt size. vLLM stops counting once it determines N + requested_output exceeds the window.
2. Exact count: your prompt contains 81921 input tokens — this is the actual token count.

The existing regex captured both indistinguishably:

r'prompt contains (?:at least )?(\d+)\s*input tokens'

This caused parse_available_output_tokens_from_error() to treat the lower bound as exact and compute a retry budget. On a retry, lowering max_tokens made the lower bound rise by the same amount, creating an infinite loop that always totals context_window + 1.

Fix

Use a negative lookahead to skip 'at least' qualified matches:

r'prompt contains (?!at least )(\d+)\s*input tokens'

• Lower-bound errors (at least N) now return None from parse_available_output_tokens_from_error() and False from is_output_cap_error(), routing correctly to input compression.
• Exact-count errors still provide a usable retry budget for output-cap reduction.

Additional changes

• Add supports_tools field to ProviderProfile for providers without function calling support (avoids sending tools to models that deadlock on unparseable schemas).
• Update test expectations to match new behavior.

Testing

Existing tests updated. test\_vllm\_token\_based\_lower\_bound\_routes\_to\_compression verifies that 'at least' messages route to compression, and test\_vllm\_exact\_count\_retry\_fits\_inside\_window confirms exact-count messages still provide a retry budget.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #58869 for the exact lower-bound vLLM parser correction. #58869 already handles this wording and adds preflight context clamping; the separate tool-support hunk here does not make the overlapping fix distinct.

[jamesreader]

Duplicate of #58869 which already covers the vLLM lower-bound fix with additional context clamping. Closing to avoid noise.