Fix CSA adaptive bank growth reference sync

[isty2e]

Summary

CSA adaptive bank growth can append entries after the reference bank has already been initialized. The previous boundary sync treated len(reference_bank.entries) == len(bank.entries) as the initialization check, so an adaptive append made the active bank longer than the reference bank and caused the reference bank to be rebuilt from the active bank.

This change gives ReferenceBank an explicit initialization state and uses that state at the boundary sync. Once the first complete reference snapshot exists, adaptive growth no longer re-sorts or rebuilds it just because the active bank grew.

Impact

• Preserves the original reference-bank snapshot across adaptive bank growth.
• Keeps clustering labels, seed-selection status, and progression masks aligned with appended bank entries.
• Keeps old checkpoint snapshots readable by deriving the initialization state when the field is absent.

Fixes #25

Validation

• uv run --python 3.11 --extra test pytest tests -q
• uv run --python 3.11 --extra test ruff check pyproject.toml src tests
• uv run --python 3.11 --extra test basedpyright src tests
• git diff --cached --check