fix: match Claude model ids in vision capability overrides

[chiruu12]

Summary

MODEL_OVERRIDES keyed vision support on claude-3 and claude-4. Anthropic moved to claude-<family>-<version> after Claude 3, so nothing is named claude-4-* and that entry matched no model id.

Every Claude model from Sonnet 4 onward therefore fell through to the binding default. On binding="custom" that default is supports_vision: False, with the comment "Per-model; set True via MODEL_OVERRIDES", so image input was reported unsupported for all of them, including AnthropicProvider.default_model (claude-sonnet-4-20250514).

Before:

claude-sonnet-4-20250514    False   (no prefix matched)
claude-opus-4-6             False   (none)
claude-opus-5               False   (none)
claude-haiku-4-5-20251001   False   (none)
claude-3-5-sonnet-20241022  True    (claude-3)

Only the legacy Claude 3 naming still matched.

Fix

Replace both version-keyed entries with a single claude- prefix. Enumerating versions is what drifted here, and every Claude model from 3 onward is multimodal. A future text-only model can be excluded with a longer, more specific key, since get_capability sorts patterns by length and the most specific one wins.

Tests

test_claude_model_ids_are_vision_capable covers the current ids across Claude 3, 4, 4.6 and 5. test_claude_override_does_not_leak_to_other_vendors pins that the broader prefix doesn't affect gpt-3.5 or gemma. The first fails on dev and passes with the change.

Ruff check, ruff format and mypy are clean on both files.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.