[Xuanwo]

Problem

test_memory_snapshot_never_drifts_from_writer_state::case_2_zero_grace_evicts_on_commit can fail at its coverage precondition even when a memtable freeze occurred. The test used the current frozen_bytes or frozen_count as evidence of a past freeze. With zero grace, a successful flush retires the frozen handle immediately, so both values can legitimately be zero after the freeze.

This was observed in the main linux-build job (run 41262655601). The drift and unflushed_bytes assertions did not fail; this change does not claim that production snapshot drift or memory corruption occurred.

Fix

Capture the initial active memtable generation and assert that it advanced after the existing 24 puts. Generation only advances when the active memtable is rotated, so it proves the freeze path ran without depending on a transient frozen handle.

After the existing flush drain:

• zero grace must report frozen_count == 0;
• a long grace must keep frozen_count > 0;
• existing frozen_bytes == 0 accounting and every assert_no_drift / unflushed_bytes invariant remain unchanged.

The change is test-only.

Validation

• Final two-case CI-profile regression: 2 passed.
• Existing test_memtable_stats_frozen_count_outlives_frozen_bytes, test_frozen_retained_during_grace_then_swept, and test_frozen_evicted_immediately_with_zero_grace: passed.
• CI-profile clippy for all targets with the workflow feature set and the Arrow ordering fixture: passed.
• Applicable pre-commit hooks (fmt, typos): passed.
• A clearly separated diagnostic experiment, which changed timing by draining before evaluating the old predicate, observed frozen_bytes=0, frozen_count=0, and generation 1 -> 3. This demonstrates the false event inference but is not an unchanged-timing reproduction and is not part of the patch.
• Hosted CI passed on this head: linux-build passed all 18,764 tests, including both cases of test_memory_snapshot_never_drifts_from_writer_state. All applicable workflows completed successfully; the path-filtered Arrow ordering compatibility job was skipped as expected.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[lance-gatekeeper]

✅ Gate recommendation: approve.

The generation check removes the timing-dependent freeze precondition while preserving the memory-accounting invariants. No blocking issues found.