plan(ml-backtesting): real-LOB integration implementation plan
10 atomic commits, ~40 sub-tasks, TDD per superpowers:writing-plans:
C1 ml-alpha loader inference_only mode
C2 ml-backtesting order types + SlotTag
C3 policy tree + bytecode flatten
C4 build.rs + book_update kernel + sim skeleton (3 fixtures)
C5 order_match kernel + latency-aware fills (4 fixtures)
C6 pnl_track kernel (1 fixture)
C7 decision_policy kernel + per-horizon ISV-Kelly (3 fixtures)
C8 BacktestHarness orchestrator + Ring 1b trainer-parity test
C9 artifacts + aggregate + fxt-backtest binary
C10 Ring 2 invariant fuzz (N in {1, 16, 256})
Plan ends with green Ring 1 (11 fixtures) + Ring 1b + Ring 2 (3 fuzz)
and a working fxt-backtest run|aggregate CLI. Ring 3 (production-day
replay), per-horizon cost-frontier sweep, model-checkpoint sweeps,
IBKR live adapter, and multi-fill P&L explicitly deferred.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
fab2d70c92