fix(models): parse Claude 3-generation version ids

[Joi Ito (Joi)]

Claude 3-generation model ids put the version before the family — claude-3-5-haiku-20241022. Every id since puts the family first — claude-haiku-4-5-20251001. _detect_version only knows the second shape.

For the first, the family word therefore sits directly against the snapshot date, and the major-only fallback rf"{family}-(\d+)(?:-|$)" matches the date:

>>> AnthropicProvider._detect_version("claude-3-5-haiku-20241022", "haiku")
(20241022, 0)
>>> AnthropicProvider._detect_version("claude-3-5-sonnet-20241022", "sonnet")
(20241022, 0)
>>> AnthropicProvider._detect_version("claude-3-7-sonnet-20250219", "sonnet")
(20250219, 0)
>>> AnthropicProvider._detect_version("claude-3-haiku-20240307", "haiku")
(20240307, 0)
>>> AnthropicProvider._detect_version("claude-3-opus-20240229", "opus")
(20240229, 0)

A major of 20241022 clears every >= gate in _get_capabilities, so these models are handed the newest capability tier.

The change: recognise the legacy shape ahead of the major-only fallback, and cap that fallback's major at two digits so no future id can have its date read as a version.

What it corrects. I diffed the full ModelCapabilities for each affected id before and after:

• Haiku 3 and 3.5 — supports_thinking and supports_native_computer_use True → False, thinking budget 32000 → 0. The capability matrix's own comment already says both are unsupported below 4.5.
• Sonnet 3.5 / 3.7 and Opus 3 — supports_1m, supports_adaptive_thinking, supports_output_config, supports_task_budget, supports_native_computer_use True → False; supported_efforts back to (low, medium, high); Opus 3 also loses speed. max_output_tokens 128000 → 64000.

What it does not correct. Two things surfaced while checking the above. I left both alone rather than widen a parser fix into a capability-matrix change, but they are worth a follow-up:

1. The 64000 output ceiling these models now land on is still too high — Claude 3-era ceilings are 4096–8192. Closer than the 128000 they had, not right.
2. The opus and sonnet branches set supports_thinking=True unconditionally, so Opus 3 and Sonnet 3.5 still advertise extended thinking. This commit does not change that in either direction.

Happy to do either as a separate PR if you want them, though the real ceilings for the retired ids may need a call from someone who can still probe them.

Why the existing tests did not catch it: the capability tests pin behavior against claude-haiku-3-5-20250929, which is not a served model id and happens to match the family-first pattern, so it parses to (3, 5) and looks correct. The new tests use the real ids.

Family-first ids are unchanged: claude-opus-4-20250514 → (4, 0), claude-sonnet-4-5-20250929 → (4, 5), claude-opus-4-8 → (4, 8), claude-fable-5 → (5, 0).

Tests: 13 added — eight parse cases across both id shapes, five capability assertions covering the corrected legacy tiers and pinning current models. Full suite 753 passed, 7 skipped.