P3-8: CSI500 independent generalization check for value/lowvol signals

[StackOverFlow11]

Summary

P3-8: does the P3-7 sign-level conclusion (value_ep/value_bp positive, volatility_20 negative on the independent 2024-07..2026-05 window) generalize outside the screened universes, specifically to CSI500 (000905.SH, mid-caps) — a cell independent in BOTH universe and time? Report-only; the unchanged P3-7 independent-validation machinery (same factor groups, same cost scenarios, same hypotheses); no new factors, no tuning, no alpha/portfolio/execution/OOS-slicing change.

• Config config/phase3_real_csi500_generalization.yaml: screened anchor SSE50|2022-2024 (must reproduce P3-6/P3-7), independent anchor SSE50|2024-2026 (must reproduce the P3-7 verdict numbers), and the NEW independent cell 000905.SH|2024-2026; CSI500|2022-2024 skipped + disclosed (runtime budget). Data feasibility probed first (monthly 500-name index_weight snapshots through 2026-05-29 incl. a pre-start 2024-06-28 snapshot).
• output.subset_report_name (the baseline_report_name precedent): each subset-validation study owns its report filename, so the P3-8 run does not clobber the accepted P3-7 artifact. Configs without the key keep the historical filename bitwise (locked by tests).
• output.subset_report_title (review HIGH fix): the report H1 is config-driven so the CSI500 study names itself ("Phase 3-8 — CSI500 Independent Generalization Check") instead of the machinery's default P3-7 phase label; the body framing stays sample-aware. P3-6/P3-7 configs leave it unset and keep the renderer's sample-aware default (regression-locked).

Real run (3 cells, ~3.55h; the 735-name CSI500 cell dominates)

• Both anchors reconciled exactly: screened raw ICs 22/22 vs the P3-5 matrix report; the independent SSE50 verdict ICs identical to P3-7 (a second reproducibility confirmation).
• CSI500 verdict: SUPPORTED (21 settled rebalances vs minimum 8): value_ep +0.0083/+0.0145, value_bp +0.0230/+0.0127, volatility_20 −0.0350/−0.0272 (train/test, both subperiods hold).
• Generalizes — and with LESS attenuation than the SSE50/CSI300 holdouts (CSI500 test |IC| 0.013–0.027 vs 0.003–0.016 there): the value/low-vol signs hold outside the screened universes; combo_ic_weighted CSI500 test IC positive in all 4 groups (highest of the three holdout cells).
• Honest caveats (disclosed): CSI500 portfolio results are positive across all groups and the full cost ladder, but legacy_trio posts +17.8%→+10.6% — another small-sample/regime ranking flip across cells; ~21 rebalances, single window. NOT a return claim.
• Secret scan: 0 occurrences in report and log; the P3-7 artifact is not overwritten.

Test plan

• pytest -p no:cacheprovider — 383 passed (+8: CSI500 config cells/groups/scenarios/hypotheses, sample classes, default + own report filename, config-driven title incl. "first line is NOT Phase 3-7", default-title regression)
• ruff check . — clean
• validate-config on all 11 configs — OK
• run-phase0 demo regression — ic 0.9600 / annual 0.8408 unchanged
• git diff --check + conflict marker scan — clean
• AGENTS.md == CLAUDE.md (verbatim copy convention)
• secret scan on the merge diff — 0 token occurrences, no artifacts/parquet committed
• real CSI500 run completed; artifact title fixed (Phase 3-8), anchors reconciled, 2/2 independent verdicts SUPPORTED. The 3.5h matrix was NOT rerun for the title fix (deterministic 1-line H1 re-render; real numbers untouched).