[BUG] analyze_market_regimes fabricates uniform probabilities and misreports method after silent fit fallback

[A1-NWS-Dev1]

🐛 Bug Description

backtesting_analyze_market_regimes can silently run entirely on the rule-based threshold classifier while still reporting method: "hmm"/"kmeans" and returning a fabricated uniform [0.333, 0.333, 0.333] probability distribution for every day in the history, regardless of the regime actually assigned.

Root cause: MarketRegimeDetector.fit_regimes has several "not enough data" early-return branches (e.g. fewer than max(50, n_regimes*20) valid feature windows -- which the tool's own default 365-day lookback usually does not clear for n_regimes=3). Each of these sets self.is_fitted = True without ever fitting self.scaler/self.model. Downstream, get_regime_probabilities then hits self.scaler.transform(...), raises NotFittedError on the never-fitted scaler, and its broad except fabricates np.ones(n) / n. Independently, detect_current_regime falls back to the rule-based threshold classifier through its own, separately-triggered exception handling -- so the regime label varies plausibly day-to-day while the probability stays fake-uniform, and self.method is never updated, so the tool keeps reporting the originally-requested method.

💰 Financial Disclaimer Acknowledgment

• I understand this is educational software and not financial advice
• I am not expecting investment recommendations or guaranteed returns
• This bug report is about technical functionality, not financial performance

📋 Reproduction Steps

1. Call backtesting_analyze_market_regimes with symbol="AAPL", method="hmm", n_regimes=3, using the tool's default lookback (no start_date/end_date).
2. Inspect the returned method field and every recent_regime_history[*].probabilities entry.

🎯 Expected Behavior

method should report what was actually used ("threshold" when the fit fell back), and probabilities should either be the fitted model's real posterior or an honest one-hot vector at the assigned regime.

Actual behavior: method still reports "hmm", and every probabilities entry is exactly [0.3333333333333333, 0.3333333333333333, 0.3333333333333333] regardless of which regime was assigned that day.

Separately: method="hmm" is also a misnomer -- it fits sklearn.mixture.GaussianMixture (static clustering, no transition-probability/temporal modeling), not a genuine Hidden Markov Model. This should be documented even where a real fit succeeds.

💻 Environment Information

• OS: Windows 11
• Python Version: 3.12.10
• MaverickMCP Version: main branch, HEAD 95c919f8ed22aa4dd0b58525900769c30a32ad3d at time of report
• Installation Method: uv sync --extra backtesting

📊 Error Messages/Logs

No exception is raised. The tool returns a successful response with a misleading method field and fabricated confidence values.

🔧 Additional Context

Confirmed by direct source inspection of maverick/backtesting/strategies/ml/regime_detector.py::MarketRegimeDetector and by empirical reproduction (instrumented calls showing NotFittedError being swallowed). A fix, accurate documentation, and regression tests are proposed in PR (linked from this issue).

🏷️ Bug Classification

Severity:

• Medium (feature partially working) -- silently fabricates analytical confidence

Component:

• MCP server/tools (backtesting domain)