feat(factors): reproduce ridge minute-return factor (PR-K, sign transfers but Reject on monotonicity + redundancy)

[StackOverFlow11]

Summary

Reproduce the ridge minute-return factor — the fifth factor of Kaiyuan Securities microstructure #27 (loop reproduction 8/10) and the only negative-sign factor in that report. This completes a fourth statistic family (RETURN) on top of the count / timing / price-level families already reproduced from the same report.

Result: Reject / Reject — but the sign transferred. IC −0.0312, ICIR −0.381 (aligned CI lower bound 0.320 > the 0.30 bar), NW-t −13.31, win rate 0.652, OOS sign-consistent (train −0.0278 / test −0.0346). Prototype (persisted pre-run, same seed-20260720 30 names as PR-I/PR-J, symbol list byte-identical): mean RankIC −0.0314, PASS. So the report's behavioural claim — ridge minutes represent retail follow-on trading whose accumulated return reflects over-reaction and therefore predicts negatively — reproduces at the IC level.

It fails on exactly two things:

1. Quantile monotonicity direction (raw +0.2 → aligned −0.2). The bucket means are Q1 0.000341 / Q2 0.000346 / Q3 0.000232 / Q4 0.000248 / Q5 0.000467 — a smile, not a reversal, with Q5 a lone outlier (final NAV 1.482 vs 1.19–1.36). The negative rank IC is real but lives in the middle of the distribution, not as a monotone tilt across sorted buckets. The reversal persists out-of-sample (test aligned −0.3), so it is not an in-sample artifact.
2. Incremental axis — orthogonalized ICIR −0.141, just under the 0.15 bar: redundant with the value/low-vol book. (Both prior PASS factors cleared this comfortably at +0.343 and +0.289.)

A clean experimental control. PR-K sits almost exactly on PR-J's machine profile — turnover 2.108 vs 2.101, half-life 8.06 vs 7.99, cross-section 472.8 vs 475.2, ridge day-validity 51.2% vs 51.5% (same mask, same scarcity gate) — which isolates the difference to the statistic, not the pipeline.

Tradability, third data point: net long-short is negative at every cost scenario (−0.001983 / −0.004092 / −0.008309), the worst of the three. Gross long-short is only +0.000125/period against turnover 2.108 — costs swamp the edge roughly 16:1 at base fee (cumulative −90.9%). aligned_spread_* is unreliable here because this factor is sign=−1 and triggers the frozen layer's cost-sign defect (sign*(gross−cost) = −gross+cost); the frozen layer was correctly NOT patched, and the warning is carried in the module docstring, config header, CLI output and run log — every place the implementer controls.

Factor ridge_minute_return_20: reuses the merged ridge mask with zero changes to the shared taxonomy (git diff main -- data/clean/intraday_volume_prv.py is empty); minute return close_t/close_{t−1}−1 with within-day lag on raw closes; daily aggregate = simple sum of selected ridge returns (pinned and test-discriminated against Π(1+r)−1, which differs >4× on the hand cases); day valid = ≥100 classifiable AND ≥10 ridge-with-return; factor = sum (not mean) of daily sums over the trailing 20 valid days; NaN below 10 valid days. Pre-registered sign −1.

Disclosed interpretation edge (task-card gap, surfaced by the implementer): the within-day lag was pinned without a 60s-adjacency requirement, so the 13:01 bar carries a ~91-minute across-lunch return summed alongside 1-minute returns. Disclosed on the spec and locked in a test; the same-slot baseline mitigates systematic selection (13:01 is compared against its own history). First parameter to vary if this factor is ever revisited.

Test plan

• Full pytest: 1527 passed (baseline 1473 + 54; the four prior factors' tests untouched, and their hand values re-asserted inside PR-K's own file) — rerun independently by orchestrator
• ruff clean; phase0 regression ic 0.9600 / annual 0.8408 unchanged
• Independent review: APPROVE, no CRITICAL/HIGH. Leak hunt clean (leakage test verified to have teeth — disabling the cutoff moves the value >50% and flips its sign, while the same late data is inert under the normal cutoff); within-day lag verified to never cross a day boundary; SUM-vs-compounding discriminator hand-recomputed and confirmed non-vacuous; frozen layer confirmed untouched.
• Secret scan: 0 token values across diff, artifacts, logs, and the tmp/ prototype records