diff --git a/docs/qc/qc-comparative-backtests.md b/docs/qc/qc-comparative-backtests.md
index 1a06807532..764ef7ff4d 100644
--- a/docs/qc/qc-comparative-backtests.md
+++ b/docs/qc/qc-comparative-backtests.md
@@ -144,6 +144,7 @@ over against that fee. Four regimes now observed:
   (frequent re-entry) are not. Batch 6 confirms: MeanReversion (low-turnover sector rotation,
   0.81→0.81, **0% delta**, PSR 46.8%) holds flat — it is multi-asset like AllWeather but, unlike
   AllWeather, rotates *within a single fee-homogeneous equity class* (<0.25 bps/trade), so its slow
+
   signals incur negligible cost. AllWeather's -30% drop came from its bonds/gold sleeve crossing
   higher per-trade friction. The discriminator is thus **signal-frequency × fee-homogeneity of the
   traded basket**, not asset-class alone.** Batch 7 sharpens this further: cadence is not turnover.
@@ -161,6 +162,7 @@ over against that fee. Four regimes now observed:
   *invalidates* the earlier broad "ML/crypto holds" generalization. AllWeather (low-turnover
   multi-asset, -30% → 0.47) and Crypto-MultiCanal (crypto indicator, -43% → 0.33) both drop
   below the robuste threshold. Crypto's 10bps fees + the indicator's signal-chasing turnover erode
+
   it hard; the Binance CASH cash-constraint benefit seen in the #1407 fee sweep (0.181→0.333)
   still leaves it well under the catalog 0.58. HAR-RV-Kelly (vol-targeting Kelly, 0.75, PSR 24.0%)
   is the exception that proves the rule: Kelly position-sizing dampens exposure, so it survives fees
@@ -463,21 +465,28 @@ ou `Sigma` est la **matrice de covariance** complete (correlations incluses). R
 10. **TrendStocks-Alpha: high return, high risk**: CAGR 15.9% mais MaxDD 39.6% (Calmar 0.40). PSR 5.8% = non significatif.
 11. **Student DualMomentum: PSR > 50%**: Sharpe 0.493 sur 2023-2025 avec MaxDD 9.0%. Seule strategie etudiante avec PSR significatif (54.9%).
 12. **Student RiskParity: performance honnete**: Sharpe 0.514, CAGR 9.3%, MaxDD 20.7%. Inverse-vol simple mais efficace. PSR 16.3% (non significatif mais respectable).
+
 13. **Student OptionWheel: catastrophe pedagogique**: Sharpe -0.51, MaxDD 103.5%. Parfait comme etude de cas du "win-rate paradoxe".
 14. **Student ValueFactor: alpha negatif confirmee**: Sharpe 0.227, PSR 0.8%. Decennie growth-dominée = facteur value sous-performant.
 15. **LeveragedETFMomentum: confirme et significatif**: Sharpe 1.779 sur 2018-2025, PSR 79.8% (2e PSR significatif apres TrendFollowing). Mais MaxDD 53.3% et CAGR 126% typiques d'un levier 3x — profil risque extreme, pas comparable aux strategies non-leveragees.
 16. **PuppiesOfTheDow et HighBookToMarketFScore: effondrement sur periode alignee**: Sharpe catalog 1.99 et 2.09 (obtenus sur leur fenetre glissante par defaut `end_date - 12 ans`) tombent a 0.302 (PSR 3.5%) et 0.411 (PSR 4.5%, MaxDD 60.4%) sur 2018-2025. Les deux meilleures lignes ML/IND du Tier 1 ne sont pas reproductibles sur la fenetre standardisee.
 17. **TrendWeather: le composite qui tient**: Sharpe 0.948 (PSR 56.6%), proche du catalog 1.16. Contraste fort avec MomentumRegime (0.185) — toutes les architectures composites ne se valent pas.
 18. **Caveat reproductibilite Trend-Following**: le code du repo backteste sur 2018-2024 donne Sharpe 0.365 / MaxDD 13.8% (backtest `3748cb62`), loin du 1.072 publie ci-dessus (`7792ae0a`, 2018-2025, etat du code cloud anterieur). Periodes differentes (2025 inclus ou non) ET drift possible repo vs cloud — a investiguer avant de citer 1.072 comme reference du code versionne.
+
 19. **MeanReversion v5.2: Best Calmar ratio**: Sharpe 0.81, MaxDD 7.5%, Calmar 1.34 — best risk-adjusted return among non-leveraged strategies. PSR 46.8% (near significance). Promoted from Tier 2 (0.29) to Tier 1. The v5.2 code (IBKR brokerage, RSI65 exit, 10% stop-loss) dramatically outperforms the older version.
 20. **AdaptiveAssetAllocation: confirmed robuste**: Sharpe 0.509, CAGR 8.0%, MaxDD 18.9% (2008-2024, 16 years). Min-var + momentum approach produces steady returns. PSR 10.6% (not significant but positive).
 21. **PairsTrading: structural failure confirmed**: Sharpe -0.28 on aligned period, PSR 0.001%. OLS hedge + cointegration still produces negative alpha. Remains exploratoire/pedagogical.
 22. **AssetClassMomentum-QC: weak aligned baseline (2026-06-22)**: 5-ETF momentum (top-3 of SPY/EFA/BND/VNQ/GSG, 252d lookback, monthly rebalance, IBKR) on 2018-2025 gives Sharpe 0.22, CAGR 6.6%, MaxDD 28.1%, PSR 3.8% (non-significant). Confirms the aligned-period momentum underperformance pattern (cf MomentumRegime 0.185, EMA-Cross-Alpha -0.010). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `6746f155`, project 33209767.
+
 23. **Cloud-RiskParity-Composite: near-flat aligned baseline (2026-06-22)**: 6-asset tactical rotation (SPY/TLT/GLD/EFA/EEM/DBC, SMA200 + 6m momentum dual filter, equal weight, monthly rebalance, IBKR) on 2018-2025 gives Sharpe 0.027, CAGR 3.5%, MaxDD 24.4%, PSR 1.2% (non-significant). Despite the "RiskParity" name the code is equal-weight momentum rotation (AQR Trend-Following style, not true risk-parity weighting). Extends the aligned-period rotation/momentum underperformance pattern (cf AssetClassMomentum 0.22 #22, Cloud-MeanReversion 0.067). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `b184b13e`, project 30820857. Note: `totalOrders=0` in the MCP wrapper is an extraction artifact (CAGR 3.5% ⇒ real trades), not a 0-trade backtest; cross-checked `list_backtests` status Completed.
 24. **Cloud-SectorRotation-Momentum: first Tier-3 (negative) aligned baseline (2026-06-22)**: 5-ETF momentum-weighted rotation (QQQ/SPY/EFA/GLD/IWM, SMA200 + 6m momentum dual filter, momentum-proportional sizing, SHY defensive, monthly rebalance, IBKR) on 2018-2025 gives Sharpe -0.029, CAGR 2.1%, MaxDD 42.7%, PSR 0.5% (non-significant). The momentum-weighted variant does *slightly worse* than the equal-weight Cloud-RiskParity-Composite #79 (-0.029 vs +0.027) — momentum-proportional sizing did not help on the aligned period. Strengthens the aligned-period momentum-underperformance finding (first baseline below zero). Promoted Tier 4 (Untested) → Tier 3 (Exploratoire). Backtest `58fb0a94`, project 30821748. Same `totalOrders=0` wrapper extraction artifact as #77/#78/#79 (CAGR 2.1% ⇒ real trades).
+
 25. **Cloud-VolTargeting: true vol-targeting, leverage-clamp raises MaxDD (2026-06-22)**: single-asset SPY vol-targeting (target_vol 12%, 21d realized-vol lookback, 30-150% allocation clamp, monthly rebalance, IBKR) on 2018-2025 gives Sharpe 0.207, CAGR 6.7%, MaxDD 38.2%, PSR 2.4% (non-significant). Genuine vol-targeting (unlike the misnomer Cloud-RiskParity-Composite #79 which is equal-weight rotation). Notable: MaxDD 38% *exceeds* AssetClassMomentum #77 (28%) — the 150% upper clamp leveraged the portfolio into vol spikes (vol↑ ⇒ target/realized demands >100% ⇒ clamped at 150% ⇒ amplified drawdown). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `179ef1c5`, project 30823587. Same totalOrders=0 wrapper artifact.
+
 26. **GlobalMacro-Regime: best backbone baseline, regime-switch + risk-parity wins (2026-06-22)**: rank-based risk-parity + SPY regime switch (Bridgewater / Antonacci 2014; SMA200+6m-mom bull/bear gate, rank×inv_vol weighting over trending risky assets in bull, inv-vol risk-parity over BND/TLT/GLD defensive in bear, monthly rebalance, IBKR) on 2018-2025 gives Sharpe **0.454**, CAGR 9.8%, MaxDD 22.8%, PSR 16.7% (respectable). The strongest aligned backbone baseline yet — nearly Tier 1 (>0.5) — and the regime-switch + inverse-vol risk-parity combination dramatically outperforms simple momentum rotations (AssetClassMomentum 0.22 #22, Cloud-RiskParity 0.027 #23, Cloud-SectorRotation -0.029 #24) while *also* keeping MaxDD controlled (22.8% vs #80 42.7%, #81 38.2%). The defensive regime rotation (risk-parity over BND/TLT/GLD in bear markets) is the likely edge over pure trend/momentum. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `bbb73b7b`, project 30781695. Same totalOrders=0 wrapper artifact.
+
 27. **MomentumRegime-AdaptiveWeights: double-defense composite destroys the SectorMomentum edge (2026-06-22)**: COMP framework composite (SectorMomentum 85% + RegimeSwitching 15% via QC Alpha / PortfolioConstruction additive; composite momentum 1/3/6/12m weights 0.5/0.2/0.2/0.1 + SMA200 filter; RegimeSwitching SPY SMA50/SMA200 bull/bear/sideways + RSI; MultiStrategyPCM groups by source_model; SPY/QQQ/IEF/GLD universe, IBKR) on 2018-2025 gives Sharpe **-0.729**, CAGR 1.875%, MaxDD 4.3%, PSR 17.4% (non-significant). The composite was *overwhelmingly defensive* on the aligned period (SMA200 filter + bear/sideways regime → mostly IEF/GLD/cash), so MaxDD is tiny (4.3%) but CAGR 1.875% < risk-free → negative Sharpe. The T85/RS15 variant is *WORSE* than the baseline 60/40 MomentumRegime composite (Sharpe 0.185) — shifting weight toward SectorMomentum added defensiveness, not edge. Confirms + extends the double-defense finding (cf Key-finding #4): standalone SectorMomentum works (Tier 1, 0.56 #22), but wrapping it in a regime-gated composite destroys the edge on the aligned period. Promoted Tier 4 (Untested) → Tier 3 (Exploratoire) — 2nd Tier-3 negative. Backtest `6e8f164f`, project 31524424. Same totalOrders=0 wrapper artifact.
+
 28. **TermStructureCommodities-QC: roll-yield signal catastrophically fails on the modern period (2026-06-22)**: long-short commodity futures on roll returns / backwardation-contango (top-quintile backwardation = long, top-quintile contango = short; near vs distant contract log-price ratio annualized by expiry gap; monthly rebalance; 21 commodity futures across Softs / Grains / Meats / Energies / Metals; IBKR margin) on 2018-2025 gives Sharpe **-0.244**, CAGR **-31.5%**, MaxDD **96.8%**, PSR 0.007% (non-significant) — the worst backbone baseline yet, nearly going to zero. This *confirms the published library header* (OOS 5Y Sharpe -0.041, MaxDD 80.8%): the roll-yield / backwardation signal that worked historically catastrophically fails on the modern period. The 2020 COVID oil crash + 2022 energy / inflation spike dislocated term structures (extreme contango then extreme backwardation), putting the strategy on the wrong long-short legs through both regimes — and the cherry-picked "Recent OOS 1.49 1Y Sharpe" sub-period in the header does not survive alignment. Reinforces the #1630 finding: published headline Sharpe figures do not survive alignment to a modern, fee-aware period. Promoted Tier 4 (Untested) → Tier 3 (Exploratoire) — 3rd Tier-3 negative (after #80 -0.029, #86 -0.729), worst by CAGR / MaxDD. Backtest `e9f7e686`, project 33224097 (created fresh; original library project 29688398 absent). Same totalOrders=0 wrapper artifact.
 
 29. **HAR-RV-J-Kelly: crypto vol-forecasting + Kelly survives alignment (2026-06-22)**: HAR-RV-J volatility forecasting (Corsi 2009 + Andersen-Bollerslev-Diebold 2007 jump component via Huang-Tauchen bipower variation; inline OLS via np.linalg.lstsq, refit every 22 days; iterated 5-day-ahead log-RV forecast; 1/4-Kelly position sizing × 5-day-momentum direction; BTC/ETH/LTC/BCH USDT on Binance) on 2018-2025 gives Sharpe **0.524**, CAGR 14.1%, MaxDD 37.1%, PSR 10.7% (non-significant). The **strongest backbone baseline since GlobalMacro-Regime (0.454, #26)**, and notably the first RISK-family baseline whose signal survives alignment with a *positive* Sharpe — contrasting sharply with TermStructureCommodities (#28, Sharpe -0.244 / MaxDD 96.8%): where the commodity roll-yield signal catastrophically broke on the modern period, the crypto jump-aware HAR-RV-J vol model + fractional Kelly held. Counter-intuitive MaxDD note: extending 2020-2025 → 2018-2025 *lowered* MaxDD (48.3% → 37.1%) — the longer OLS coefficient path through the 2022 crypto bear produced a less-leveraged exposure than the shorter window. PSR 10.7% = non-significant (crypto single-directional long-only + 1/4-Kelly cap keeps it conservative). Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `df687834`, project 31650567. Same totalOrders=0 wrapper extraction artifact (CAGR 14.1% ⇒ real trades).
@@ -490,19 +499,31 @@ ou `Sigma` est la **matrice de covariance** complete (correlations incluses). R
 
 33. **composite-c1-multiasset: COMP framework ensemble adds little over single momentum, alpha/PCM choice > wiring (2026-06-23)**: QC Alpha Framework C4.1 Multi-Asset Rotation composite (3-model alpha ensemble: MomentumAlpha 12m-1m risk-adjusted + MACDAlpha crossover + RelativeStrengthAlpha cross-asset 3m; RiskParityPCM equal-weight weekly 100% max-exposure 40% sector-cap; DrawdownCap 12% + trailing-stop 4%; VWAP 4-slice execution; universe SPY/TLT/GLD/USO/EFA at 2x leverage, IBKR margin) on 2018-2025 gives Sharpe **0.258**, CAGR 6.490%, MaxDD 17.0%, PSR 8.7% (non-significant). A weak positive Sharpe — the 3-alpha ensemble + framework architecture adds **little over a single momentum signal** on the aligned period (cf AssetClassMomentum 0.22 #22, Vol-Ensemble-Conservative 0.265 #31). MaxDD is controlled at 17% by the 12% drawdown cap (vs 22.8% for the uncapped GlobalMacro-Regime #26). It sits **far below the framework composite leaders** FamaFrenchAllWeather (0.684) and EMATrend (0.741): **COMP composites vary widely, so the component alpha/PCM choice matters more than the framework wiring itself** — the same Alpha Framework scaffold produces a 0.258 weakling here vs 0.684-0.741 leaders depending on which alphas and PCM are plugged in. Notably the aligned Sharpe 0.258 is *higher* than the catalog-campaign figure (0.175), so alignment did not degrade this composite. PSR 8.7% non-significant. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `1236cf59`, project 32981093 (pre-existing; 5 framework files verified present, only main.py date-aligned). Same totalOrders=0 wrapper extraction artifact (CAGR 6.49% ⇒ real trades).
 
-34. **composite-c2-equityfactor: strongest COMP backbone — fine-fundamental stock factor investing dominates static ETF rotation (2026-06-23)**: QC Alpha Framework C4.2 Equity Factor composite (4-model alpha ensemble: ValueAlpha P/E+P/B, QualityAlpha ROE+debt-to-equity, LowVolAlpha realized-vol, MomentumFactorAlpha price momentum; MeanVariancePCM weekly 65% max-exposure 18% sector-cap; SectorCapRiskModel 10% sector weight + 0.8 beta + trailing 4% stop; TWAP 6-slice execution; FineFundamentalUniverseSelectionModel coarse top-200 by dollar-volume then fine top-25 by market cap, IBKR margin) on 2018-2025 gives Sharpe **0.574**, CAGR 11.942%, MaxDD 18.6%, PSR 25.8% (non-significant but the highest PSR among the no-ML backbones). The **strongest COMP (composite-framework) backbone verified to date** — near Tier-1 (>0.5) — and it **holds and slightly improves** over the catalog window (0.543 → 0.574, PSR 16.8% → 25.8%), genuinely robust rather than period-overfit. The defining contrast is with the sibling **composite-c1-multiasset (Sharpe 0.258, #33)**: both share the *same* Alpha Framework scaffold, yet c2 more than doubles c1's Sharpe. The difference is the **universe and factors, not the wiring**: c2 invests in individual large-cap US stocks selected by fine fundamentals and scored across Value/Quality/LowVol/Momentum (true factor investing across 25 single-name equities with a mean-variance PCM), whereas c1 rotates 5 static asset-class ETFs on a momentum/MACD/relative-strength ensemble. Within the COMP family this confirms and sharpens the #33 finding — **the component alpha/PCM choice and the stock-level fundamental universe matter far more than the framework wiring itself**, and a fine-fundamental stock universe is where the Alpha Framework actually earns its Sharpe. It remains below the framework leaders FamaFrenchAllWeather (0.684) and EMATrend (0.741), but closes most of the gap c1 left open. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `8eecba32`, project 32981222 (pre-existing catalog project `c2-equityfactor-post2801` bt `3173f8b39`; only main.py date-aligned). Same totalOrders=0 wrapper extraction artifact (CAGR 11.9% ⇒ real trades).
+34. **composite-c2-equityfactor: strongest COMP backbone — fine-fundamental stock factor investing dominates static ETF rotation (2026-06-23)**: QC Alpha Framework C4.2 Equity Factor composite (4-model alpha ensemble: ValueAlpha P/E+P/B, QualityAlpha ROE+debt-to-equity, LowVolAlpha realized-vol, MomentumFactorAlpha price momentum; MeanVariancePCM weekly 65% max-exposure 18% sector-cap; SectorCapRiskModel 10% sector weight + 0.8 beta + trailing 4% stop; TWAP 6-slice execution; FineFundamentalUniverseSelectionModel coarse top-200 by dollar-volume then fine top-25 by market cap, IBKR margin) on 2018-2025 gives Sharpe **0.574**, CAGR 11.942%, MaxDD 18.6%, PSR 25.8% (non-significant but the highest PSR among the no-ML backbones). The **strongest COMP (composite-framework) backbone verified to date** — near Tier-1 (>0.5) — and it **holds and slightly improves** over the catalog window (0.543 → 0.574, PSR 16.8% → 25.8%), genuinely robust rather than period-overfit. The defining contrast is with the sibling **composite-c1-multiasset (Sharpe 0.258, #33)**: both share the *same* Alpha Framework scaffold, yet c2 more than doubles c1's Sharpe. The difference is the **universe and factors, not the wiring**: c2 invests in individual large-cap US stocks selected by fine fundamentals and scored across Value/Quality/LowVol/Momentum (true factor investing across 25 single-name equities with a mean-variance PCM), whereas c1 rotates 5 static asset-class ETFs on a momentum/MACD/relative-strength ensemble. 
+
+Within the COMP family this confirms and sharpens the #33 finding — **the component alpha/PCM choice and the stock-level fundamental universe matter far more than the framework wiring itself**, and a fine-fundamental stock universe is where the Alpha Framework actually earns its Sharpe. It remains below the framework leaders FamaFrenchAllWeather (0.684) and EMATrend (0.741), but closes most of the gap c1 left open. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `8eecba32`, project 32981222 (pre-existing catalog project `c2-equityfactor-post2801` bt `3173f8b39`; only main.py date-aligned). Same totalOrders=0 wrapper extraction artifact (CAGR 11.9% ⇒ real trades).
+
+35. **FamaFrenchAllWeather: headline collapses on alignment — the 0.684 / PSR 87.5% OOS figure was a small-sample artifact (2026-06-23)**: QC Alpha Framework composite FamaFrench (20%) + AllWeather (80%) (FamaFrenchAlpha risk-adjusted-momentum top-2 rotation over VLUE/MTUM/SIZE/QUAL/USMV, skip-month, quarterly, NO SMA200 filter — AllWeather handles defense; AllWeatherAlpha static SPY/IEF/GLD/XLP Ray-Dalio-inspired monthly drift-rebalance; MultiStrategyPCM 20/80 allocation monthly; NullRiskManagement + ImmediateExecution; IBKR margin) on 2018-2025 gives Sharpe **0.338**, CAGR 6.578%, MaxDD 13.1%, PSR 22.9% (non-significant). **A clear period-overfitting collapse**: the headline figures do *not* survive alignment. The README's 2010-2026 sweep Sharpe 0.588 falls to 0.338 (-42%), and — most tellingly — a separately-run OOS 2023-2026 backtest (`b08c8956`) that reported Sharpe 0.684 / PSR **87.5%** on only **835 tradeable dates** collapses to 0.338 / PSR 22.9% on the full 1761-date aligned window: the 87.5% PSR was a **small-OOS-sample artifact**, not a robust statistical edge. The same-config catalog FF20/AW80 over 2015-2025 (`70415edc`) = 0.472. On the aligned window the 80% AllWeather sleeve (mostly static SPY/IEF/GLD/XLP) drags the composite *below* the standalone FamaFrench rotation (0.445, #32) — though it buys tight drawdown control (MaxDD 13.1%, among the tighter backbones alongside Vol-GARCH 10.8% #30 / Vol-Ensemble 10.4% #31). 
+
+**This corrects the "FamaFrenchAllWeather 0.684 framework leader" references in findings #33 and #34** — that 0.684 was the short OOS window, not the aligned baseline; the real aligned COMP leaderboard is composite-c2-equityfactor 0.574 (#34), not FamaFrenchAllWeather. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `e9ac7c66`, project 28882145 (pre-existing; cloud main.py was an OOS 2023-2026 variant, replaced with the canonical repo version date-aligned to 2018-2025). Same totalOrders=0 wrapper extraction artifact (CAGR 6.6% ⇒ real trades).
 
-35. **FamaFrenchAllWeather: headline collapses on alignment — the 0.684 / PSR 87.5% OOS figure was a small-sample artifact (2026-06-23)**: QC Alpha Framework composite FamaFrench (20%) + AllWeather (80%) (FamaFrenchAlpha risk-adjusted-momentum top-2 rotation over VLUE/MTUM/SIZE/QUAL/USMV, skip-month, quarterly, NO SMA200 filter — AllWeather handles defense; AllWeatherAlpha static SPY/IEF/GLD/XLP Ray-Dalio-inspired monthly drift-rebalance; MultiStrategyPCM 20/80 allocation monthly; NullRiskManagement + ImmediateExecution; IBKR margin) on 2018-2025 gives Sharpe **0.338**, CAGR 6.578%, MaxDD 13.1%, PSR 22.9% (non-significant). **A clear period-overfitting collapse**: the headline figures do *not* survive alignment. The README's 2010-2026 sweep Sharpe 0.588 falls to 0.338 (-42%), and — most tellingly — a separately-run OOS 2023-2026 backtest (`b08c8956`) that reported Sharpe 0.684 / PSR **87.5%** on only **835 tradeable dates** collapses to 0.338 / PSR 22.9% on the full 1761-date aligned window: the 87.5% PSR was a **small-OOS-sample artifact**, not a robust statistical edge. The same-config catalog FF20/AW80 over 2015-2025 (`70415edc`) = 0.472. On the aligned window the 80% AllWeather sleeve (mostly static SPY/IEF/GLD/XLP) drags the composite *below* the standalone FamaFrench rotation (0.445, #32) — though it buys tight drawdown control (MaxDD 13.1%, among the tighter backbones alongside Vol-GARCH 10.8% #30 / Vol-Ensemble 10.4% #31). **This corrects the "FamaFrenchAllWeather 0.684 framework leader" references in findings #33 and #34** — that 0.684 was the short OOS window, not the aligned baseline; the real aligned COMP leaderboard is composite-c2-equityfactor 0.574 (#34), not FamaFrenchAllWeather. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `e9ac7c66`, project 28882145 (pre-existing; cloud main.py was an OOS 2023-2026 variant, replaced with the canonical repo version date-aligned to 2018-2025). Same totalOrders=0 wrapper extraction artifact (CAGR 6.6% ⇒ real trades).
+36. **EMATrend: survives alignment (0.741 → 0.611) — the highest-Sharpe COMP backbone, but Mag7-concentrated (survivorship caveat) (2026-06-23)**: QC Alpha Framework composite EMA-Cross (70%) + TrendStocks (30%) (EMACrossAlpha 20/50 EMA on the 5 Mag7 AAPL/MSFT/GOOGL/AMZN/NVDA, daily emission; TrendStocksAlpha double-confirmation Price>SMA200 + EMA20>EMA50 on 15 mega-caps including the same 5 Mag7, weekly; MultiStrategyPCM 70/30 weekly; NullRiskManagement + ImmediateExecution; IBKR margin) on 2018-2025 gives Sharpe **0.611**, CAGR 16.670%, MaxDD 27.9%, PSR 19.8% (non-significant). **Survives alignment with a mild drop** (catalog 0.741 @ 2015-2025 → 0.611, -18%) — NOT a period-overfit collapse, in direct contrast to the sibling framework composite FamaFrenchAllWeather (#35, 0.588→0.338). The drop is the expected cost of losing the 2015-2017 Mag7 pre-ramp and absorbing the 2022 Mag7 drawdown; the trend signal itself holds. On the aligned window this is the **highest-Sharpe COMP backbone verified to date**, edging composite-c2-equityfactor 0.574 (#34), and it carries the highest CAGR (16.67%) of any no-ML backbone. **But the Mag7 survivorship caveat dominates the interpretation**: the EMA sleeve is 100% Mag7, so a meaningful fraction of the Sharpe is an artifact of the Mag7 outperformance regime that defined the 2015-2025 decade rather than a transferable trend-following edge. Highest Sharpe is **not** the same as the most robust constitution — composite-c2-equityfactor (0.574, fine-fundamental factor investing across 25 large-cap stocks) is the more defensible COMP leader constitution-wise; EMATrend is the higher-Sharpe but Mag7-concentrated one. 
 
-36. **EMATrend: survives alignment (0.741 → 0.611) — the highest-Sharpe COMP backbone, but Mag7-concentrated (survivorship caveat) (2026-06-23)**: QC Alpha Framework composite EMA-Cross (70%) + TrendStocks (30%) (EMACrossAlpha 20/50 EMA on the 5 Mag7 AAPL/MSFT/GOOGL/AMZN/NVDA, daily emission; TrendStocksAlpha double-confirmation Price>SMA200 + EMA20>EMA50 on 15 mega-caps including the same 5 Mag7, weekly; MultiStrategyPCM 70/30 weekly; NullRiskManagement + ImmediateExecution; IBKR margin) on 2018-2025 gives Sharpe **0.611**, CAGR 16.670%, MaxDD 27.9%, PSR 19.8% (non-significant). **Survives alignment with a mild drop** (catalog 0.741 @ 2015-2025 → 0.611, -18%) — NOT a period-overfit collapse, in direct contrast to the sibling framework composite FamaFrenchAllWeather (#35, 0.588→0.338). The drop is the expected cost of losing the 2015-2017 Mag7 pre-ramp and absorbing the 2022 Mag7 drawdown; the trend signal itself holds. On the aligned window this is the **highest-Sharpe COMP backbone verified to date**, edging composite-c2-equityfactor 0.574 (#34), and it carries the highest CAGR (16.67%) of any no-ML backbone. **But the Mag7 survivorship caveat dominates the interpretation**: the EMA sleeve is 100% Mag7, so a meaningful fraction of the Sharpe is an artifact of the Mag7 outperformance regime that defined the 2015-2025 decade rather than a transferable trend-following edge. Highest Sharpe is **not** the same as the most robust constitution — composite-c2-equityfactor (0.574, fine-fundamental factor investing across 25 large-cap stocks) is the more defensible COMP leader constitution-wise; EMATrend is the higher-Sharpe but Mag7-concentrated one. G.1 note: the repo `main.py` had drifted to the EMA40/Trend60 sweep *starting point* (its own docstring), while the deployed QC Cloud project (28911253) and the catalog entry both use the 70/30 sweep *winner* — the repo was converged to the winner so the aligned run is apples-to-apples vs the catalog 0.741. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `3095a263d5bd30df181ec002c0a52b72`, project 28911253. Same totalOrders=0 wrapper extraction artifact (CAGR 16.7% ⇒ real trades).
+G.1 note: the repo `main.py` had drifted to the EMA40/Trend60 sweep *starting point* (its own docstring), while the deployed QC Cloud project (28911253) and the catalog entry both use the 70/30 sweep *winner* — the repo was converged to the winner so the aligned run is apples-to-apples vs the catalog 0.741. Promoted Tier 4 (Untested) → Tier 2 (Historique). Backtest `3095a263d5bd30df181ec002c0a52b72`, project 28911253. Same totalOrders=0 wrapper extraction artifact (CAGR 16.7% ⇒ real trades).
 
 37. **Research-Executor is a research harness (not a strategy) — the #1630 no-ML backbone campaign is complete (2026-06-23)**: a G.1 verification of the queued "#1630 next floor = #92 Research-Executor (dernier leader COMP à vérifier)" against the source reveals Research-Executor is a **research execution harness, not a tradable strategy**. Its `main.py` sets a 2-day window (2024-01-02 → 2024-01-03), runs 8 embedded research notebooks via a `MockQB` shim, writes the executed notebooks to the object store, and calls `self.quit('Done')` inside `initialize` (with `on_data = pass`) — there is no portfolio, no PnL and no tradable mechanics, so an "aligned 2018-2025 baseline" is meaningless (the algorithm quits on day 1 regardless of the date range). Its README and `projects/catalog.json` already classify it `Type: Utility (research execution harness, not a trading strategy)` / `classification: untested`; it was **miscategorized** in the deep-queue as a COMP leader. This **closes the #1630 no-ML backbone campaign**: every IND/COMP/RISK/FACTOR candidate (#72, #77–#91) is verified across Key-findings #22–#36 (15 baselines), with the aligned-COMP leaderboard led by EMATrend 0.611 (#36, highest Sharpe, Mag7 survivorship caveat) and composite-c2-equityfactor 0.574 (#34, the most robust constitution). The remaining Tier-4 entries are characterized in the table note above (OPT naked-options with a MaxDD > 100 % simulator caveat; ML/DL/RL deferred to multi-seed #6).
 
-38. **Adaptive-Conformal-Risk: real ACI overlay survives real fees (low-turnover immunity), but the Sharpe is Mag7 survivorship, not conformal value (2026-06-23)**: Adaptive Conformal Inference risk overlay on multi-factor momentum (ACI algorithm, Gibbs & Candès 2021: online alpha-adjustment `alpha_{t+1} = alpha_t + gamma·(1{violation} − target_alpha)` over a 60-day rolling nonconformity-score window, empirical-quantile prediction interval with finite-sample `1+1/√n` correction; position sizing `signal·confidence/interval_width` — wider conformal interval → smaller position; 3-window momentum signal 21/63/126d, sector cap 30%, target-vol 15%, monthly rebalance; 15 large-caps across 5 sectors, IBKR margin; source: ECE student project El Bakkali Gr02, Issue #238) on 2018-2025 gives Sharpe **0.449**, CAGR 11.525%, MaxDD 22.5%, PSR 12.526% (non-significant), 2011 tradeable dates. This is the **first real #1630-aligned vetting** of a previously-unvetted Tier-2 ML strategy: the catalog 0.42 was unverified (it matches no Cloud backtest), and the only pre-existing Cloud run ("ACI-Risk-v1", project 29841071, Sharpe 0.604 / 2015-2026) had **silently lost its `set_brokerage_model(IBKR)` line** in the deployed copy and ran on the negligible default fee model — i.e. not #1630-compliant. Two findings, one positive and one a caveat: **(a) the low-turnover fee-resistance pattern holds** — the aligned-with-IBKR 0.449 sits only modestly below the no-fee 0.604, and much of that gap is the harder 2018-2025 window (no 2015-2017 bull-momentum tailwind) rather than pure fee drag; the ACI vol-targeting actively *scales exposure down*, cutting turnover, which *helps* fee-resistance. A genuinely sophisticated conformal-prediction risk overlay does **not** collapse under real fees — reinforcing the #1630 realized-turnover discriminator (cf SectorMomentum 0.56 #22, BlackLitterman 0.83 near-immune): it is turnover that kills strategies, not mechanism sophistication. **(b) But the Mag7-survivorship caveat dominates the interpretation** — the universe is AAPL/MSFT/NVDA/AMZN/TSLA (5 of 7 Mag7) plus 10 financials/healthcare/consumer/industrial large-caps, and over 2018-2025 the Mag7 (NVDA especially) ran enormously, so long-only momentum concentrates in them and the 0.449 is substantially survivorship drift rather than ACI-overlay marginal value (the ACI vol-targeting may even *drag* vs uncapped Mag7 momentum). The overlay's contribution cannot be isolated without a uniform-sizing control — same caveat class as EMATrend (#36, Mag7 sleeve). Net: a real, fee-resistant SOTA mechanism deployed on a survivorship-biased universe → the honest headline is "0.449, but it is mostly Mag7, not conformal prediction." Promoted Tier 2 (Historique, unvetted —) → Tier 2 (Historique, verified). Backtest `cce5af0271e5e5989908897b6c5ecb08`, project 33278416 (baseline-clone `1630-baseline-AdaptiveConformalRisk`, created fresh; only main.py date-aligned to 2018-2025 + IBKR brokerage restored to match the canonical local copy). Same totalOrders=0 wrapper extraction artifact (CAGR 11.5% ⇒ real trades).
+38. **Adaptive-Conformal-Risk: real ACI overlay survives real fees (low-turnover immunity), but the Sharpe is Mag7 survivorship, not conformal value (2026-06-23)**: Adaptive Conformal Inference risk overlay on multi-factor momentum (ACI algorithm, Gibbs & Candès 2021: online alpha-adjustment `alpha_{t+1} = alpha_t + gamma·(1{violation} − target_alpha)` over a 60-day rolling nonconformity-score window, empirical-quantile prediction interval with finite-sample `1+1/√n` correction; position sizing `signal·confidence/interval_width` — wider conformal interval → smaller position; 3-window momentum signal 21/63/126d, sector cap 30%, target-vol 15%, monthly rebalance; 15 large-caps across 5 sectors, IBKR margin; source: ECE student project El Bakkali Gr02, Issue #238) on 2018-2025 gives Sharpe **0.449**, CAGR 11.525%, MaxDD 22.5%, PSR 12.526% (non-significant), 2011 tradeable dates. This is the **first real #1630-aligned vetting** of a previously-unvetted Tier-2 ML strategy: the catalog 0.42 was unverified (it matches no Cloud backtest), and the only pre-existing Cloud run ("ACI-Risk-v1", project 29841071, Sharpe 0.604 / 2015-2026) had **silently lost its `set_brokerage_model(IBKR)` line** in the deployed copy and ran on the negligible default fee model — i.e. not #1630-compliant. Two findings, one positive and one a caveat: **(a) the low-turnover fee-resistance pattern holds** — the aligned-with-IBKR 0.449 sits only modestly below the no-fee 0.604, and much of that gap is the harder 2018-2025 window (no 2015-2017 bull-momentum tailwind) rather than pure fee drag; the ACI vol-targeting actively *scales exposure down*, cutting turnover, which *helps* fee-resistance. 
 
-39. **composite-c2-equityfactor: regime-robustness split — the 0.574 is bull-market-concentrated (pro-cyclical long-only), not regime-agnostic; capital-preserving but unprofitable in the 2022 bear (2026-06-23)**: a 4-regime sub-period split of the strongest COMP backbone (composite-c2-equityfactor, full-window Sharpe 0.574 / 1761 tradeable dates, #34) on an IBKR-fee clone (`1630-robustness-c2-regime`, project 33280244; clone main.py = canonical + date parameterization only, all other logic byte-identical) decomposes the aligned 2018-2024 window into four market regimes — R1 2018-19 pre-COVID normal, R2 2020-21 COVID bull, R3 2022 bear/rate-hike, R4 2023-24 recovery/AI. The split partitions the window **exactly** (R1 503 + R2 505 + R3 251 + R4 502 = 1761 = full window, no overlap/gap). **Result — sharply bimodal, NOT uniform**: R2 COVID-bull **Sharpe 1.040 / CAGR 18.9% / PSR 50.5%** and R4 recovery **Sharpe 1.094 / CAGR 22.6% / PSR 82.7%** are where c2 earns *all* its edge (both statistically meaningful; both US-large-cap bull/recovery markets — the 2020-21 stimulus rally and the 2023-24 AI/Mag7 rally); R1 pre-COVID normal is weak (**Sharpe 0.131 / PSR 17.1%**) and R3 2022 bear is **negative (Sharpe −0.219 / CAGR −0.7%)** — c2 *loses money* in the rate-hike bear. The full-window 0.574 is thus a blend dominated by the two bull/recovery regimes (~57% of dates at Sharpe ~1.07) offsetting the weak/negative normal+bear regimes (~43% at ~0.0). **This nuances finding #34's "genuinely robust rather than period-overfit"**: c2 is indeed robust to *fee/window alignment* (the #1630 discriminator — it holds at 0.574 with real IBKR fees, doesn't collapse on alignment like FamaFrenchAllWeather #35), but it is **NOT regime-agnostic** — it is a pro-cyclical long-only US-large-cap factor strategy whose Sharpe is a bull-market premium, exactly as expected for its design (Value/Quality/LowVol/Momentum on FineFundamental top-25 market-cap). The defensible nuance: c2's risk model (SectorCapRiskModel 18% portfolio-DD circuit breaker + 4% trailing stops) **does protect capital in the bear** — the 2022 MaxDD is only 9.5% (well under the 18% cap), so c2 *survives* downturns without profiting from them, a meaningful property for a long-only backbone. **Contrast with EMATrend (#36)**: EMATrend's caveat is *survivorship* (Mag7-universe bias); c2's caveat here is *regime-concentration* (pro-cyclical) — but c2's *constitution* is cleaner (factor-diversified across 25 stocks, no Mag7 concentration), so the "most robust COMP constitution" claim (#34) stands *constitutionally* even as the Sharpe is regime-concentrated. **Practical takeaway**: do not over-read the 0.574 as a transferable all-weather edge — it is a large-cap-US bull/recovery premium with capital preservation in bears, not a regime-agnostic alpha. Backtests R1 `63605ad6` / R2 `7fb12884` / R3 `1342b6b5` / R4 `7dc742af` (clone project 33280244, compile `614c4d79`; canonical c2 project 32981222 / full-window bt `8eecba32` unchanged). Same totalOrders=0 wrapper extraction artifact (regime CAGRs 4.8%/18.9%/−0.7%/22.6% ⇒ real trades). `See #1630`.
+A genuinely sophisticated conformal-prediction risk overlay does **not** collapse under real fees — reinforcing the #1630 realized-turnover discriminator (cf SectorMomentum 0.56 #22, BlackLitterman 0.83 near-immune): it is turnover that kills strategies, not mechanism sophistication. **(b) But the Mag7-survivorship caveat dominates the interpretation** — the universe is AAPL/MSFT/NVDA/AMZN/TSLA (5 of 7 Mag7) plus 10 financials/healthcare/consumer/industrial large-caps, and over 2018-2025 the Mag7 (NVDA especially) ran enormously, so long-only momentum concentrates in them and the 0.449 is substantially survivorship drift rather than ACI-overlay marginal value (the ACI vol-targeting may even *drag* vs uncapped Mag7 momentum). The overlay's contribution cannot be isolated without a uniform-sizing control — same caveat class as EMATrend (#36, Mag7 sleeve). Net: a real, fee-resistant SOTA mechanism deployed on a survivorship-biased universe → the honest headline is "0.449, but it is mostly Mag7, not conformal prediction." Promoted Tier 2 (Historique, unvetted —) → Tier 2 (Historique, verified). Backtest `cce5af0271e5e5989908897b6c5ecb08`, project 33278416 (baseline-clone `1630-baseline-AdaptiveConformalRisk`, created fresh; only main.py date-aligned to 2018-2025 + IBKR brokerage restored to match the canonical local copy). Same totalOrders=0 wrapper extraction artifact (CAGR 11.5% ⇒ real trades).
 
-40. **PCA-StatArbitrage: catalog 0.40 does not survive alignment (0.399 → 0.205, PSR 1.4% noise) — and the drop is a WINDOW-effect, not fee (fee-IMMUNITY proven by a hardcoded no-brokerage clone) (2026-06-23)**: PCA statistical-arbitrage mean-reversion (sklearn PCA + per-stock OLS on log prices → residual z-scores → contrarian long the z<−1.5 tail, weights ∝ z-deviation, full monthly rotation via `set_holdings(liquidate=True)`; top-100 US equities by dollar volume, coarse-universe monthly refresh, IBKR margin; source: Hands-On AI Trading Ch.06 Ex.13 / Avellaneda-Lee 2010) on 2018-2025 gives Sharpe **0.205**, CAGR 6.762%, MaxDD 31.8%, PSR **1.418%** (noise-level), 2011 tradeable dates. The **first real #1630-aligned vetting** of a previously-unvetted Tier-2 ML strategy: catalog 0.40 (README 0.399) does not survive alignment — a −49% drop, PSR 1.4% = indistinguishable from noise. Two findings, both correcting a pre-run prediction: **(a) the drop is a WINDOW-effect, NOT a fee-effect** — the catalog README 0.399 was already an IBKR number (the canonical `main.py` carries `set_brokerage_model(IBKR)`), so the 0.399→0.205 degradation cannot be fee-remediation; it is the aligned window's 2024-25 Mag7-momentum melt-up, structurally hostile to contrarian mean-reversion (the "oversold" residual names keep falling, the winners keep winning, so the z<−1.5 contrarian tail bleeds). **(b) Fee-IMMUNE — refuting the "monthly full-rotation ⇒ fee-vulnerable" prediction** — a hardcoded no-brokerage clone of the identical logic (project 33281920, backtest `63fce57d`) returns byte-identical Sharpe 0.205 / CAGR 6.762% / MaxDD 31.8% / PSR 1.418%; physically removing the brokerage changes nothing at 3-decimal precision. The pre-run hypothesis was that the monthly full-rotation of a *changing* contrarian subset would make this the archetypal fee-collapse case; the result instead **extends the #1630 discriminator regime-3 near-immune class to PCA stat-arb** — the fee-HOMOGENEITY of the US-equity top-100 basket (<0.25 bps/trade via IBKR, cf EMA-Cross-Stocks IBKR 0.991 = no-brokerage 0.991, fee-finding #2 at line 740) dominates over the per-event turnover size, exactly as for EMA-CS / ML-RandomForest / composite-c2 (#34). The discriminator is basket fee-homogeneity × realized turnover, not mechanism (mean-reversion vs trend) or rotation frequency in isolation. Net honest headline: "PCA-StatArb = 0.205 on alignment, PSR 1.4% noise, and it is the aligned WINDOW that broke it, not the fees." Promoted Tier 2 (Historique, unvetted —) → Tier 2 (Historique, verified, un-robust). Backtest `9169c962` (IBKR) / `63fce57d` (no-fee control), project 33281920 (baseline-clone `1630-baseline-PCAStatArbitrage`, created fresh; only main.py date-aligned to 2018-2025, IBKR unchanged = catalog standard). Same totalOrders=0 wrapper extraction artifact (CAGR 6.76% ⇒ real trades). Numbered #40 assuming #4089 (c2-regime) takes #39 on rebase post-#4084-merge.
+39. **composite-c2-equityfactor: regime-robustness split — the 0.574 is bull-market-concentrated (pro-cyclical long-only), not regime-agnostic; capital-preserving but unprofitable in the 2022 bear (2026-06-23)**: a 4-regime sub-period split of the strongest COMP backbone (composite-c2-equityfactor, full-window Sharpe 0.574 / 1761 tradeable dates, #34) on an IBKR-fee clone (`1630-robustness-c2-regime`, project 33280244; clone main.py = canonical + date parameterization only, all other logic byte-identical) decomposes the aligned 2018-2024 window into four market regimes — R1 2018-19 pre-COVID normal, R2 2020-21 COVID bull, R3 2022 bear/rate-hike, R4 2023-24 recovery/AI. The split partitions the window **exactly** (R1 503 + R2 505 + R3 251 + R4 502 = 1761 = full window, no overlap/gap). **Result — sharply bimodal, NOT uniform**: R2 COVID-bull **Sharpe 1.040 / CAGR 18.9% / PSR 50.5%** and R4 recovery **Sharpe 1.094 / CAGR 22.6% / PSR 82.7%** are where c2 earns *all* its edge (both statistically meaningful; both US-large-cap bull/recovery markets — the 2020-21 stimulus rally and the 2023-24 AI/Mag7 rally); R1 pre-COVID normal is weak (**Sharpe 0.131 / PSR 17.1%**) and R3 2022 bear is **negative (Sharpe −0.219 / CAGR −0.7%)** — c2 *loses money* in the rate-hike bear. The full-window 0.574 is thus a blend dominated by the two bull/recovery regimes (~57% of dates at Sharpe ~1.07) offsetting the weak/negative normal+bear regimes (~43% at ~0.0). 
+
+**This nuances finding #34's "genuinely robust rather than period-overfit"**: c2 is indeed robust to *fee/window alignment* (the #1630 discriminator — it holds at 0.574 with real IBKR fees, doesn't collapse on alignment like FamaFrenchAllWeather #35), but it is **NOT regime-agnostic** — it is a pro-cyclical long-only US-large-cap factor strategy whose Sharpe is a bull-market premium, exactly as expected for its design (Value/Quality/LowVol/Momentum on FineFundamental top-25 market-cap). The defensible nuance: c2's risk model (SectorCapRiskModel 18% portfolio-DD circuit breaker + 4% trailing stops) **does protect capital in the bear** — the 2022 MaxDD is only 9.5% (well under the 18% cap), so c2 *survives* downturns without profiting from them, a meaningful property for a long-only backbone. **Contrast with EMATrend (#36)**: EMATrend's caveat is *survivorship* (Mag7-universe bias); c2's caveat here is *regime-concentration* (pro-cyclical) — but c2's *constitution* is cleaner (factor-diversified across 25 stocks, no Mag7 concentration), so the "most robust COMP constitution" claim (#34) stands *constitutionally* even as the Sharpe is regime-concentrated. **Practical takeaway**: do not over-read the 0.574 as a transferable all-weather edge — it is a large-cap-US bull/recovery premium with capital preservation in bears, not a regime-agnostic alpha. Backtests R1 `63605ad6` / R2 `7fb12884` / R3 `1342b6b5` / R4 `7dc742af` (clone project 33280244, compile `614c4d79`; canonical c2 project 32981222 / full-window bt `8eecba32` unchanged). Same totalOrders=0 wrapper extraction artifact (regime CAGRs 4.8%/18.9%/−0.7%/22.6% ⇒ real trades). `See #1630`.
+
+40. **PCA-StatArbitrage: catalog 0.40 does not survive alignment (0.399 → 0.205, PSR 1.4% noise) — and the drop is a WINDOW-effect, not fee (fee-IMMUNITY proven by a hardcoded no-brokerage clone) (2026-06-23)**: PCA statistical-arbitrage mean-reversion (sklearn PCA + per-stock OLS on log prices → residual z-scores → contrarian long the z<−1.5 tail, weights ∝ z-deviation, full monthly rotation via `set_holdings(liquidate=True)`; top-100 US equities by dollar volume, coarse-universe monthly refresh, IBKR margin; source: Hands-On AI Trading Ch.06 Ex.13 / Avellaneda-Lee 2010) on 2018-2025 gives Sharpe **0.205**, CAGR 6.762%, MaxDD 31.8%, PSR **1.418%** (noise-level), 2011 tradeable dates. The **first real #1630-aligned vetting** of a previously-unvetted Tier-2 ML strategy: catalog 0.40 (README 0.399) does not survive alignment — a −49% drop, PSR 1.4% = indistinguishable from noise. Two findings, both correcting a pre-run prediction: **(a) the drop is a WINDOW-effect, NOT a fee-effect** — the catalog README 0.399 was already an IBKR number (the canonical `main.py` carries `set_brokerage_model(IBKR)`), so the 0.399→0.205 degradation cannot be fee-remediation; it is the aligned window's 2024-25 Mag7-momentum melt-up, structurally hostile to contrarian mean-reversion (the "oversold" residual names keep falling, the winners keep winning, so the z<−1.5 contrarian tail bleeds). 
+
+**(b) Fee-IMMUNE — refuting the "monthly full-rotation ⇒ fee-vulnerable" prediction** — a hardcoded no-brokerage clone of the identical logic (project 33281920, backtest `63fce57d`) returns byte-identical Sharpe 0.205 / CAGR 6.762% / MaxDD 31.8% / PSR 1.418%; physically removing the brokerage changes nothing at 3-decimal precision. The pre-run hypothesis was that the monthly full-rotation of a *changing* contrarian subset would make this the archetypal fee-collapse case; the result instead **extends the #1630 discriminator regime-3 near-immune class to PCA stat-arb** — the fee-HOMOGENEITY of the US-equity top-100 basket (<0.25 bps/trade via IBKR, cf EMA-Cross-Stocks IBKR 0.991 = no-brokerage 0.991, fee-finding #2 at line 740) dominates over the per-event turnover size, exactly as for EMA-CS / ML-RandomForest / composite-c2 (#34). The discriminator is basket fee-homogeneity × realized turnover, not mechanism (mean-reversion vs trend) or rotation frequency in isolation. Net honest headline: "PCA-StatArb = 0.205 on alignment, PSR 1.4% noise, and it is the aligned WINDOW that broke it, not the fees." Promoted Tier 2 (Historique, unvetted —) → Tier 2 (Historique, verified, un-robust). Backtest `9169c962` (IBKR) / `63fce57d` (no-fee control), project 33281920 (baseline-clone `1630-baseline-PCAStatArbitrage`, created fresh; only main.py date-aligned to 2018-2025, IBKR unchanged = catalog standard). Same totalOrders=0 wrapper extraction artifact (CAGR 6.76% ⇒ real trades). Numbered #40 assuming #4089 (c2-regime) takes #39 on rebase post-#4084-merge.
 
 ## #1630 Aligned Backbone Leaderboard (2018-2025, no-ML)
 
@@ -582,6 +603,7 @@ The table makes the family hierarchy explicit: the regime switch (#26) is the di
 5. ~~**Transaction cost re-backtest**: Add `SetBrokerageModel` + configurable brokerage parameter~~ — Done, #2575 + fee sweep EMA-Cross-Stocks + Crypto-MultiCanal (See #2471, #2575, #2588)
 6. **Cross-seed validation (gates #7)**: ≥4 seeds (0/1/7/42/99) for ML/DL/RL strategies — multi-cycle, requires re-training each model on ≥4 seeds and re-backtesting (heavy QC + GPU). **Blocks #7**: σ_cross_seed is a prerequisite input for Edge vs σ. See section "Edge vs σ — statut & dépendance" below.
 7. **Edge vs σ (gated on #6)**: Compute `(Sharpe - baseline_Sharpe) / σ_cross_seed` for ML/DL/RL multi-seed strategies vs B&H baseline. **Not computable standalone** — requires σ_cross_seed from #6. For non-ML (IND/COMP/RISK/OPT) strategies there is no σ_cross_seed (single-run), so Edge vs σ applies only to the ML/DL/RL subset once #6 delivers. See section "Edge vs σ — statut & dépendance" below.
+
 8. **Trend-Following repo/cloud drift**: repo code gives Sharpe 0.365 on 2018-2024 vs published 1.072 (2018-2025, prior cloud state) — identify which code version produced 1.072 and align repo (see Key finding 18)
 9. **No-ML backbone campaign complete (2026-06-23)**: all IND/COMP/RISK/FACTOR Tier-4 candidates verified (Key-findings #22–#36, 15 baselines). Remaining Tier-4 = OPT (#73/#75, naked-options MaxDD > 100 % simulator caveat) + ML/DL/RL (#63–#71/#76, deferred to #6 multi-seed / training-specialist). Research-Executor (#92) is a research harness, not a strategy — out of baseline scope (Key-finding #37).