C1a — Lookahead and recursive bias detection

[eaitbrahim]

PRD §3 C1. Freqtrade ships lookahead-analysis and recursive-analysis; keel has no equivalent.

Why this is not covered by PBO/CSCV

They catch different failure modes. PBO/CSCV asks "did I overfit by selecting among many configs?" Lookahead bias asks "does this strategy use information it could not have had?" A strategy with a single config, never swept, can still be catastrophically lookahead-biased — and PBO would report nothing.

This is the failure that makes a backtest look brilliant and a live run lose money.

What Freqtrade does

• lookahead-analysis — re-runs backtests on truncated data slices and diffs the signals. If a signal on bar N changes when bars after N are removed, the strategy is reading the future.
• recursive-analysis — detects indicators whose values change depending on how much history is fed in (recursive/non-deterministic formulas), a lighter check.

Fit with keel

keel's engine already has the right seam: Rule.detect(candles_by_tf) -> Setup | None is documented as a pure decision with no side effects (keel/strategy/rules/base.py:106-157). Purity is exactly what makes truncation-diffing meaningful — call detect() on a truncated series and on the full series and compare.

Worth checking against the engine's higher-timeframe bearish-bias veto (keel/strategy/engine.py:196-213), which reads "the coarsest available higher timeframe" — a natural place for leakage if the coarse bar is not yet closed.

Scope

• keel rules lookahead <rule-id> and a recursive-indicator variant.
• Report which bar/indicator diverged, not just pass/fail.
• Decide whether this becomes a promotion-gate input (see the PRD's open question 1). Recommendation: yes — a rule that fails lookahead analysis should not be promotable, which is precisely the enforcement posture no competitor has.