feat(factors): P3-1 multi-factor pipeline + first real price+financial baseline

[StackOverFlow11]

Summary

P3-1 — first REAL multi-factor baseline (price + PIT financials). The pipeline previously consumed only the FIRST enabled factor; this PR makes it consume every enabled factor and adds the first real multi-factor baseline: momentum_20 + roe + netprofit_yoy on SSE50, still combined by the same equal-weight alpha. No parameter search, no learned weights, not a performance claim — this validates the multi-factor plumbing against real data.

What changed

• _build_factors (was _build_factor): instantiates every enabled factor in config order; duplicate factor names are a readable config error. Single-factor configs are byte-stable (primary = the only factor).
• Batched financials: _maybe_enrich_financials collects ALL financial-factor fields and does one fina_indicator fetch + one asof_financials pass — each field independently honours ann_date <= trade_date; no per-factor refetch. Demo + financial factor still raises a readable error (no fabricated financials).
• Multi-column factor panel: one column per factor; processing (drop_missing → neutralize → zscore) was already per-column. drop_missing requires all enabled factors per name/date — disclosed in DOWNGRADES, never scored on partial data.
• Analytics: _factor_analytics now returns per-factor {ic_mean, ic_ir, quantile_returns, coverage} plus combo-score analytics (forward returns computed once). Top-level metrics remain the PRIMARY (first-enabled) factor's → identical for all existing configs. The alphalens cross-check stays on the primary raw factor (report-only, P2-4 contract intact).
• Report: active factor list; per-factor coverage/IC/quantile tables + combo row; financial ann_date coverage per field, labelled TRADED financial factor vs diagnostic only. output.baseline_report_name keeps the phase3 report file separate from phase2's.
• Config: config/phase3_real_multifactor.yaml (SSE50, 2023-07 ~ 2024-06 — same window as phase2 for comparability; neutralize PIT SW-L1; top_n 20). Runs via the shared run-phase2-baseline real-baseline runner.

Real-data results (manual, not in CI)

Regression — phase2 single-factor rerun (13.6 min): UNCHANGED
annual −10.19% · IC 0.0083 · PIT SW-L1 coverage 98.53% — identical to the P2-3 baseline. Demo also unchanged (ic 0.96 / annual 0.84).

Phase3 multi-factor (14.3 min):

Factor	coverage	IC mean	IC IR
momentum_20	91.70%	0.0083	0.0272
roe	100.00%	0.0006	0.0033
netprofit_yoy	100.00%	0.0001	0.0006
combo score (equal-weight)	—	−0.0038	−0.0279

• Portfolio: annual −9.05% (single-factor was −10.19%) · maxDD −16.17% · sharpe −0.59 · avg turnover 0.85
• Financial ann_date coverage 100% on both fields (reported per field, labelled TRADED)
• momentum_20 IC identical across the two runs (0.0083) — cross-run consistency evidence for the per-factor path
• Financial-factor ICs ≈ 0 on this 50-name / 1-year window — disclosed honestly; this is plumbing validation, not a signal claim
• Standard analytics both ran (quantstats CAGR −14.38% / alphalens IC 0.0083 == simple IC)

Quality gates

Gate	Result
pytest -p no:cacheprovider	249 passed (was 234: +10 multifactor, +1 multi-field as-of, +4 report surface)
ruff check .	clean
validate-config ×4 (demo / example_tushare / phase2 / phase3)	OK
run-phase0 demo	ic 0.96 / annual 0.84 unchanged
run-phase2-baseline (phase2 config)	−10.19% / 0.0083 unchanged
run-phase2-baseline (phase3 config)	new baseline above
Hygiene	no token, artifacts/parquet git-ignored, no conflict markers

Tests (network-free)

test_multifactor_pipeline.py (10): factor list build (order / disabled / duplicate / none), single-fetch batched financials + as-of correctness + input immutability, demo+financial readable error, e2e demo multi-factor panel + per-factor/combo analytics + primary==first, report factor list + combo + no-secret, single-factor legacy shape. test_pit_financials.py (+1): multiple fields as-of aligned in one call. test_phase2_baseline.py (+4): active-factor list, per-field role labels, per-factor table, configurable baseline report name.

Docs

RUNBOOK (Phase 3-1 section), TEST_REPORT (counts + P3-1 notes), BIAS_AUDIT (multi-factor ann_date / combination disclosure, regenerated from renderer), CLAUDE.md / AGENTS.md progress.