feat(neutralize): P2-3 point-in-time SW industry covariate (replace current-tag)

[StackOverFlow11]

Summary

P2-3 upgrades the industry neutralization covariate from the CURRENT stock_basic.industry tag (broadcast to every date — a membership look-ahead) to a point-in-time, as-of-trade-date SW industry, with the SW level configurable (processing.neutralize.industry_level, L1/L2/L3, default L1). No new factor, no parameter tuning; the framework layering and the neutralize math are unchanged (it already drops NaN-industry rows).

What changed

• data/feed/tushare_covariates.py — pit_sw_intervals(symbols, level) reads tushare index_member_all for each symbol's SW membership intervals at the chosen level (l1/l2/l3_name, in_date/out_date; out_date None = active). The industry() current-tag accessor is retained but no longer wired into neutralization.
• data/clean/pit_industry.py (new) — asof_industry(index, intervals) picks, per (trade_date, symbol), the industry whose [in_date, out_date) covers that date (new industry wins on a reclassification date — PIT-safe; pre-start membership carries forward; no covering interval / unknown symbol → NaN).
• data/clean/covariates.py — enrich_pit_industry(panel, industry_series) sets a per-date industry column.
• qt/config.py — processing.neutralize.industry_level (L1/L2/L3, default L1).
• qt/pipeline.py — _maybe_enrich_covariates builds the PIT industry at the configured level; logs coverage. A name with no SW history stays NaN — a disclosed coverage gap the neutralizer drops — never a silent current-tag fallback. _collect_downgrades rewritten (PIT, configurable level, UNI-010).
• qt/phase2_baseline.py + qt/reports.py — the phase2 report discloses the actual SW level + PIT industry coverage.
• Docs — BIAS_AUDIT.md (regenerated), RUNBOOK.md, TEST_REPORT.md, CLAUDE.md, AGENTS.md.

Real-data finding (this is the important part)

Going PIT is not behaviour-neutral, and a real L1-vs-L2 comparison disproved the initial "L2 isolates granularity" hypothesis. On the SSE50 baseline (2023-07..2024-06, momentum_20):

Industry source	buckets	annual return	Sharpe
old stock_basic.industry (current tag)	~100 (tushare's own)	−17.6%	−0.96
PIT SW-L1	31	−10.2%	−0.57
PIT SW-L2	~130	−9.3%	−0.50

(Raw-factor IC is unchanged at 0.0083 — neutralization affects the portfolio via the processed scores, not the raw IC.)

L1 ≈ L2, so the −17.6 → −10 jump is the tushare → SW taxonomy switch, which is inherent to going point-in-time: only the SW classification carries the in_date/out_date history needed for an as-of alignment; the old stock_basic.industry tag cannot be PIT-aligned at all. It is not a granularity choice.

Default = SW-L1 (31 broad sectors): the standard granularity for industry neutralization and the DOF-safest on small cross-sections (~50 names/cross-section on SSE50; L2's ~130 sub-industries approach saturation). L2/L3 remain available via config.

Real baseline (SSE50, manual, not in CI)

Default SW-L1: run-phase2-baseline OK — 68 symbols, 11 settled rebalances, PIT SW-L1 coverage 98.53% (67/68; 1 symbol has no SW history → NaN, disclosed), annual return −10.2%. The report regenerates with the actual level + coverage line.

Quality gates

Gate	Result
pytest -p no:cacheprovider	223 passed, 0 failed (P0=95 / P1=79 / P2-1=15 / P2-2=22 / P2-3=12)
ruff check .	clean
validate-config × (example / example_tushare / phase2_real_baseline)	all OK
run-phase0 (demo)	OK (P0/P1 numbers unchanged — neutralization is off on demo)

Hygiene: no tushare token, conflict markers, or artifacts//*.parquet tracked.

Tests (network-free)

test_pit_industry.py (12): as-of switch / carry-forward / missing→NaN / overlap, enrich_pit_industry, pipeline wiring (per-date industry, configured level passed, no current-tag fallback), and industry_level config (default L1, accepts L1/L2/L3, rejects invalid, phase2 config = L1). + level-selective feed parse and report-level disclosure in the covariates / real-path / bias-audit / phase2 tests.