fix(llm): keep label retries above token floor

[Synvoya]

Fixes #2086.

Root cause

The adaptive community-label retry path reduces max_tokens as it bisects a failed batch. The earlier #1690 adjustment improved the formula from 64 + 24*n to 256 + 48*n, but a single-community retry on current v8 still receives only 304 output tokens. That leaves the recovery base case with the smallest budget.

Change

• Preserve the existing batch-size scaling and 8192-token cap.
• Enforce a 512-token minimum before passing the value through _resolve_max_tokens().
• Keep GRAPHIFY_MAX_OUTPUT_TOKENS authoritative when users explicitly set it.
• Add a regression test that forces a two-community parse failure, follows both single-community retries, and asserts that every call receives the 512-token floor.

Validation

• Regression test before fix: failed with budgets [352, 304, 304].
• python -m pytest tests/test_label_retry.py -q: 2 passed.
• Full locked Python 3.12 suite: 3550 passed, 3 skipped.
• ruff check graphify/llm.py tests/test_label_retry.py: passed.
• All five tools.skillgen validation modes: passed.
• graphify update .: completed successfully.
• git diff --check: passed.