[brndnmtthws]

Defect

With hard_band_rebalance_fraction < 1 (e.g. 0.5), a hard-band step sizes each order as round(gap * fraction). Python rounds half to even, so a one-share gap gives round(-0.5) == 0. A position whose target weight is zero therefore keeps its last share forever. Because a zero target with a live position counts as full drift (relative_drift = 1.0), that share sets off a hard rebalance on every run, which keeps re-sizing (and churning) the rest of the basket.

Live evidence

Seen in a live regime-rebalance account during a long zero-target (risk-off) spell for one symbol: the exit halved each run (… → 123 → 61 → 30 → 15 → 7 → 3 → 1) and then stopped at one share. After that, every run was a hard rebalance, about 5.6 orders per session on average, until the symbol's target turned positive again. A replay of 2018–2026 showed no measurable CAGR/drawdown effect, so this fixes order hygiene and a stuck state. It is not a returns change.

Fix

Round the step magnitude up, with a 1e-9 tolerance so float products such as 10 * 0.3 don't round up a whole share:

desired = int(math.copysign(math.ceil(abs(desired) * rebalance_fraction - 1e-9), desired))

Every hard step now moves at least one share toward the target and never past it (ceil(|gap| * f) <= |gap| for f <= 1). Soft-band, flow and deficit paths are unchanged, a full step (fraction == 1.0) is unchanged, and exits stay staged (492 → 246 → 123 → 61 → 30 → 15 → 7 → 3 → 1 → 0).

Behavioral impact: odd partial gaps now trade one share more than before (e.g. 3 of a 5-share gap instead of 2). Risk controls are unaffected.

Tests

New tests in tests/test_regime_rebalance.py:

• zero target with one share sells it on the hard band
• 5-share gaps at fraction 0.5 trade 3 and never overshoot
• a 492-share zero-target exit finishes at 0 instead of stranding one share

All three fail on main (077e457) and pass with the fix. Existing partial-rebalance tests use even gaps and are unchanged.

Validation run locally:

• uv run pytest: 821 passed
• uv run ruff check ., uv run ruff format --check ., uv run ty check: clean

Merge caution

Deployments that track the main image pick this up on their next scheduled run. It changes live order sizing, so merge deliberately.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-16T20:44:46.802970Z	f1e3d86	Draft marked ready
🔒 Security Review	✅ Completed 2026-10-16T20:45:26.265099Z	f1e3d86	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.