feat: I5d MMP quintile 5Y group backtest (exploratory, stacked on I5c)

[StackOverFlow11]

I5d — MMP Quintile 5Y Group Backtest (EXPLORATORY)

This is an EXPLORATORY grouped factor analysis of the I5c MMP minute-derived
daily score — NOT a performance claim and NOT parameter tuning.

Stacked on I5c (PR #37). This PR is based on feat/i5c-mmp-minute-factor
(PR #37, still open) and its diff shows ONLY the I5d commits. Merge/rebase after
I5c lands; the base will be retargeted to main once #37 merges. Dependency is
disclosed, not hidden.

What it does

On each monthly rebalance date the cross-section is ranked by the PIT-safe daily
score intraday_mmp20_ew_0930_1450 (the I5c Minute Microstructure Pressure
factor, visible at the 14:50 cutoff) and split into analytics.quantiles=5
EQUAL-COUNT rank buckets (Q1 = lowest score, Q5 = highest). Each group runs as
its own long-only equal-weight portfolio through the SAME event-driven machinery —
a FRESH BacktestEngine + IntradayTailEventModel + SimExecution(fee_rate=0.001)
per group — with I5b raw stk_limit execution feasibility ON (14:50 decision /
14:51 execution, exec-to-exec returns, NEVER close-to-close).

Design highlights

• Memory-conscious minute load: the monthly model only needs minute bars on
  the ANCHOR dates (rebalance ∪ exit), so the runner reads ONLY those days from
  the persistent intraday cache, one trading day at a time. Cache-only, read-only:
  stk_mins_live_calls=0 (a miss never warms/backfills). 12.9M raw rows over a
  5-year CSI500 window without materialising the full minute history.
• One shared exec-price matrix (immutable) across the 5 fresh per-group models
  via a new backward-compatible precomputed_prices param on
  IntradayTailEventModel (default None = byte-identical I5a/I5b build).
• Equal-count RANK buckets (not value cuts): deterministic on ties / degenerate
  scores and robust to the MMP factor's fat tails.

Result (one window — exploratory)

CSI500 (000905.SH), 2021-06-01 → 2026-05-31, 59 monthly rebalances, 892/995
constituents minute-covered (103 uncovered dropped + disclosed; window NOT
shortened).

group	final NAV	annual	Sharpe	maxDD
Q1 (low)	0.9822	-0.36%	0.085	-39.1%
Q2	1.0423	0.85%	0.142	-36.6%
Q3	1.0495	0.99%	0.149	-38.3%
Q4	1.1275	2.47%	0.217	-34.4%
Q5 (high)	1.1577	3.02%	0.240	-37.2%

Final NAV and annual return increase monotonically Q1→Q5 (group monotonicity
Spearman = 1.0000); synthetic Q5−Q1 spread +0.29%/period, +17.15% cumulative.
Higher MMP (more intraday buying pressure) → higher realized next-month return
over this window. ⚠️ One overlapping window, one universe, no robustness, no
tuning — far too little to infer factor quality. The Q5−Q1 spread is a synthetic
long-only leg difference, NOT a separately executed dollar-neutral book.

Invariants preserved

• Daily close-to-close, BacktestEngine, SimExecution, the MMP factor math and
  the I5b execution feasibility are UNCHANGED — only grouping + per-group
  orchestration is new.
• raw stk_limit feasibility ON (strict coverage 52584/52584); per-group blocked
  counts do not double-count across groups; missing bar blocks (never daily-close
  fallback).

Tests & gates

• pytest 577 passed (+19 I5d); ruff clean; all 17 configs validate;
  run-phase0 ic_mean=0.9600 / annual_return=0.8408 unchanged; I5c short smoke
  final_nav=0.957143 unchanged (zero drift from the precomputed_prices change).
• Report + logs + figures secret-scanned: 0 token leaks.