Fix future-price leakage in indicator aggregate fallbacks

[Aarav-Nagar]

Problem

TechnicalIndicators.calculate_indicators receives the complete backtest frame. Its short-history and library-failure fallbacks broadcast the full frame's close mean/min/max to earlier rows, allowing later prices to change earlier indicators.

For closes [100, 110, 90], the SMA fallback was [100, 100, 100]; at the second bar, only 100 and 110 are available, so the fallback should be 105.

Change

• Use expanding means and extrema through each row for aggregate fallbacks, including the exception handler.
• Preserve normal pandas-ta calculations, their warm-up NaNs, and existing neutral defaults. These running values remain fallback approximations, not fully initialized indicators.
• Add 52 synthetic regression cases covering future-price perturbations, indicator-length boundaries, missing library results/columns, early and late library exceptions, explicit expected fallback values, and unchanged normal library outputs.
• Document reproduction and scope in docs/indicator-lookahead-regression.md.

The test approach is inspired by Freqtrade lookahead-analysis. This is original ATL test code, not a port of Freqtrade's CLI: timestamps and frame length are fixed, future prices are perturbed, and earlier indicator values must remain exactly equal.

Validation

Against original features.py at 85c02d5f09670b3a39405346a5d4b17f78413ff2, the new tests produce 49 failed, 3 passed. With the fix, all 52 are included in this passing run:

python -m pytest dashboard/backend/tests/backtesting dashboard/backend/tests/test_market_data_store.py dashboard/backend/tests/domain/backtesting/test_external_run_service_move.py --timeout=180 -q --tb=short

277 passed, 2 skipped locally on Windows/Python 3.12 with the pinned project dependencies. pip check and git diff --check pass. The full backend suite is left to repository CI.

This does not prove platform-wide absence of lookahead or restart equivalence, and no portfolio-return improvement is claimed. Extending a frame across the existing minimum-history thresholds can still switch fallback values to library warm-up NaNs; a uniform readiness policy is a separate change.

[vercel]

@Aarav-Nagar is attempting to deploy a commit to the allan-feng's projects Team on Vercel.

A member of the Team first needs to authorize it.