Fix provider_for for full Claude model IDs

[askhatsoltanov1984-lang]

Problem

ModelRegistry.provider_for(...) previously matched only the short Claude
aliases (opus / sonnet / haiku). Clients that pass canonical Claude
model IDs such as:

• claude-opus-4-7
• claude-sonnet-4-6
• claude-haiku-4-5-20251001

fell through the Claude branch, then the Gemini branch (no prefix match),
and ended up classified as codex. Downstream this caused Claude-bound
sessions to be routed to the Codex CLI, which silently broke session
continuity for clients that pass the full model name.

Fix

Also match a claude- prefix in provider_for. Short aliases continue
to work unchanged; non-Claude IDs (gpt-*, gemini-*, o3, etc.) are
untouched.

if model_id in CLAUDE_MODELS or model_id.startswith("claude-"):
    return "claude"

Tests

Added tests/test_provider_for_claude_full_ids.py covering:

• Full Claude IDs (claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5-20251001) -> claude
• Existing short aliases (opus, sonnet, haiku) -> still claude
• Non-Claude IDs (gpt-5.2-codex, gemini-2.5-pro) -> unchanged

Pure unit test — no network, no API calls. Runs in the existing pytest
suite.

Why merge

The patch is logic-only, additive, and matches the existing pattern
already used for Gemini (prefix check alongside a known-set check). The
short-alias path is unchanged so no in-tree behaviour regresses.