Baseline pairing is now “no excuses” in CI/release

Release: InvarLock 0.3.4 — Plus lower-memory retries for HF adapters

Highlights

• CI/Release baseline pairing now fails closed with strict parity checks.
• Report generation requires paired-windows evidence (non-perfect pairing is rejected).
• Chunked snapshot/restore helps avoid peak-memory spikes during retries.

This is one of those releases that makes the framework more “serious” in the places where it should be. If you say you’re paired to a baseline in CI/release mode, you now have to prove it with the right evidence, and the system enforces parity (dataset/tokenizer/masking) instead of letting ambiguity slip through.

It’s a stricter posture, but it’s also the kind of strictness that prevents painful downstream debates. When a report says it’s paired, it really is paired.

On the practical side, chunked snapshot/restore helps reduce peak memory during retries—useful when you’re operating close to the edge on bigger models or tighter nodes.

For more details, see CHANGELOG.md.