fix(llm): probe Windows Claude CLI candidates before spawning

[DebadityaHait]

Summary

On Windows, Graphify currently trusts the first claude.cmd found on PATH. A stale or broken npm shim can shadow a working native Claude CLI and make both Claude CLI call paths fail.

This is a resolver-only split of #2573. The all-batches-failed artifact write guard remains separate in #2726.

Changes

• centralize Windows Claude CLI resolution for extraction and lightweight labeling;
• probe claude.cmd, claude.exe, then claude with a bounded --version call;
• skip candidates that return non-zero or raise an OS/subprocess error;
• pass the selected path to the real -p invocation;
• preserve non-Windows bare claude behavior;
• add mocked regression coverage for fallback and all-candidate failure.

Validation

• uv run --frozen pytest tests/test_claude_cli_backend.py -q — 38 passed
• uv run --frozen pytest tests/test_claude_cli_backend.py tests/test_labeling.py -q — 65 passed
• uv run --frozen pytest tests/test_llm_backends.py tests/test_extract.py tests/test_extract_cli.py -q --tb=short — 287 passed, 8 skipped
• uv run --frozen ruff check graphify/llm.py tests/test_claude_cli_backend.py — passed
• python -m py_compile graphify/llm.py tests/test_claude_cli_backend.py — passed
• uv run --frozen graphify update . — completed
• git diff --check — passed

Related to #2573.