Fix operator-precedence bug in sparse L0 tracking condition

[MaxGhenis]

Summary

Finding #4 (HIGH). i % tracking_n / 2 == 0 parsed as (i % tracking_n) / 2 == 0, which is equivalent to i % tracking_n == 0 because any non-zero remainder divided by 2 is non-zero. The intended 2x logging density (the sparse loop runs 2x as many epochs as the dense loop) was silently lost, halving the fidelity of the sparse performance DataFrame that feeds the dashboard.

Stride is now max(1, tracking_n // 2) with a comment explaining both the intent and the original parse mistake.

Test plan

• Add tests/test_sparse_tracking.py verifying the sparse DataFrame has ~4x the dense DataFrame row count at 100 dense epochs (2x density × 2x epoch count).
• All existing tests pass (uv run pytest tests -x -q -> 16 passed).

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
microcalibrate	Ready	Preview, Comment	Apr 17, 2026 0:40am

[MaxGhenis]

LGTM (cannot self-approve; posting as comment).

• sparse_tracking_n = max(1, tracking_n // 2) gives the intended 2x log density for the sparse loop (which runs 2x as many epochs). The inline comment explains the parse mistake ((i % tracking_n) / 2 == i % tracking_n == 0) so the intent is obvious on future reads.
• max(1, ...) guards tracking_n == 1 (when epochs <= 10), preventing a divide-by-zero on i % 0.
• Test exercises the L0 path via csv_path and reads the sparse CSV back, counting rows to verify the ~4x ratio (2x density × 2x sparse epochs). Tight regression.

Docstring / code comment explains what the tracking density should be (match the 2x epoch ratio). Intent is now self-documenting.