Phase 9 Session 1 - EDGAR XBRL fundamental data layer + margin compression alpha factor

[shiviancodes]

Summary

• New EDGAR XBRL companyfacts data layer: migration 0011 (form_type, accession_number, UNIQUE), nexus/data/edgar/forms/form_xbrl.py parser with filed >= end integrity
  invariant, and scripts/ingest_fundamentals.py per-ticker idempotent ingest. 74,662 rows / 137 of 140 tickers covered (3 foreign 20-F filers absent as expected).
• Four fundamental factors in nexus/signals/factors/fundamental.py with pure helpers + DB-backed wrappers; opt-in dispatch wiring in backtest.py mirrors the
  supply_panel pattern.
• fundamental_margin_compression (t = +4.834 at 126d) is the first NEXUS factor to pass HLZ Bonferroni / Holm / BH at M=400. Composer sign-flipped vs published
  literature — the empirical direction at 126d on the semiconductor universe is margin compression → outperformance (mean reversion / priced-for-perfection). CALM-regime t =
  +6.35; sub-window late-third t = +5.13 (strengthening, not decaying).
• fundamental_roa (full-window t = +3.76) intentionally NOT registered — late-third decay to t = -0.06 with 6/6 sign split mirrors the Phase 5 momentum decay pattern.
  Documented for revisit after 2026 annual filings.
• rd_intensity and asset_growth null at all three horizons.
• Rolling registry refreshed: fundamental_margin_compression mean_IC positive in every window (min +0.016, max +0.181) so the aggregator mean_ic > 0 gate accepts it at
  every rebalance.
• Paper trader Phase 9 vs Phase 8 baseline: CAGR +8.72% (vs +5.97%), Sharpe 0.488 (vs 0.374), Max DD -32.68% (vs -36.68%). The previously-dead 2025-09-30 → 2026-03-31
  100%-cash window from Phase 8 is no longer dead.
• 150/150 tests pass (excluding test_health.py which needs live FastAPI).