feat(research): I5e MMP quintile CSI300 independent generalization (EXPLORATORY)

[StackOverFlow11]

Summary

Freeze the accepted I5d MMP quintile grouped-backtest machinery and run an
independent-universe generalization check for the MMP intraday signal. I5d (PR
#39) found clean Q1→Q5 monotonicity on CSI500 over one 5-year window; that is
not enough evidence. This PR re-runs the SAME frozen machinery on CSI300
(000300.SH) over the SAME window (2021-06-01 → 2026-05-31) so monotonicity can be
attributed to the factor rather than the original universe/regime.

Config-only change — the MMP math, equal-count bucket semantics (Q1 = lowest,
Q5 = highest), the I5a/I5b intraday tail event model (14:50 decision / 14:51
execution, exec-to-exec returns), raw stk_limit execution feasibility, and the
cache-only anchor-date minute loading are all unchanged. Report name/title are
already config-driven, so no code change is needed. EXPLORATORY, not a
performance claim; no tuning / no learned alpha / no new factors.

Changed files (1)

• config/phase_i5e_mmp_quintile_csi300_generalization.yaml (new) — CSI300 cell,
  unique I5e project/output/report names + title; everything else identical to the
  accepted I5d config.

Preflight (independent cell choice)

• CSI300 minute-cache coverage over the full window: 448 / 481 = 93.1% covered
  (≥ accepted CSI500 89.6%) → enough cross-section for 5 equal-count groups.
• No fallback to CSI1000 needed; the 5-year window is NOT shortened.
• CSI300 index_weight was only cached for ~1 recent year (ledger honest, no false
  coverage); the missing membership was fetched via the existing P4 daily
  read-through cache (21 index_weight calls) — exactly as I5d did. Minute bars
  stay cache-only.

Result — monotonicity FAILS to generalize

covered=448/481, stk_mins_live_calls=0, rebalances=59, fee_rate=0.001

group	final NAV	annual	Sharpe	maxDD
Q1 (low MMP)	0.8809	-2.55%	-0.04	-34.3%
Q2	0.7377	-6.00%	-0.26	-38.7%
Q3	0.9418	-1.21%	+0.02	-34.2%
Q4	0.7573	-5.50%	-0.20	-44.2%
Q5 (high MMP)	0.6665	-7.92%	-0.30	-53.9%

• monotonicity Spearman: −0.5000 (final NAV & annual) — vs +1.0000 on CSI500.
• Q5−Q1 synthetic spread: mean −0.45%/period, cumulative −26.77% (vs +17.15% on CSI500).
• Q5 (high MMP) is the WORST group, not the best. High-MMP groups do not
  outperform low-MMP groups on CSI300.

Verdict: FAILS — the I5d CSI500 monotonicity does NOT generalize to CSI300 (it
reverses), strongly suggesting the CSI500 result was universe/regime-specific. This
is exactly the negative evidence the independent check was designed to surface. The
CSI500 comparison is shown as context only, never as a performance claim.

Report / log / figures (gitignored artifacts):
artifacts/reports/phase_i5e_mmp_quintile_csi300_generalization.md (+ _figures/),
artifacts/logs/phase_i5e_mmp_quintile_csi300_generalization.log.

Gates (all green)

• targeted I5a/I5b/I5c/I5d tests: pass (machinery frozen)
• full pytest: 678 passed (677 baseline + 1 from the test_cache_config.py
  config-glob parametrization auto-covering the new config)
• ruff: clean · validate-config: 18/18 OK · run-phase0: ic_mean=0.9600,
  annual_return=0.8408 unchanged
• git diff --check clean; no merge markers
• secret scan: real token value 0 hits in config / report / log / figures;
  report/log contain no .config.json / tushare.token literals

Non-goals honored

No MMP/bucket/event-timing/execution-feasibility change; no learned alpha / IC
weights / new factors / tuning; no partial-fill/liquidity realism; no schema
registry / D6; no live data-update or minute backfill (stk_mins_live_calls=0);
no progress-doc edits (AGENTS.md / CLAUDE.md / contract doc untouched — separate
post-acceptance goal); PR #38 untouched.

⚠️ Do not merge until Codex acceptance.