Add Binance L2 virtual-MBO reconstruction study

[tradingexpert]

Summary

• add a deterministic Binance L2 + individual-trade to virtual-MBO reconstruction model
• expose named queue-conservative and queue-optimistic policies instead of presenting inferred queues as observed FIFO
• add a streaming full-capture study with exchange-time alignment, sequence checks, top-of-book validation, and JSON manifests
• document the assumptions, evidence, recommended default, and segment boundaries across the public docs
• make the crypto execution path prominent in the README and publish a measurable realism challenge in issue Crypto realism challenge: validate L2 fill models against execution evidence #62

Full-capture evidence

The study used the completed BTCUSDT and ETHUSDT capture through 2026-08-10T17:12:57Z, before the machine-wide network outage.

Measure	BTCUSDT	ETHUSDT
Snapshot-anchored segments	4	3
Depth endpoints reconstructed	1,668,215	1,659,917
Individual trades aligned	2,847,061	4,498,386
Broken segments	0	0
Late trades with validated buffer	0	0
Exact book-ticker joins	44,277 / 44,277	72,104 / 72,104
Required replenishment / traded quantity	10.90%	14.65%

Receive ordering was measured rather than assumed. A 2-second buffer left 13,422 ETH trades late and a 10-second buffer left 71; the 60-second default left none. This is an offline evidence-alignment window, not simulated exchange latency.

Zero-price, zero-quantity @trade messages remain preserved in raw capture, are counted in the report, and are excluded from executable MBO flow. No raw capture data or generated study reports are committed.

Validation

• ruff check .
• 28 focused tests passed, including reconstructed-stream Python/C++ equivalence
• per-module branch coverage: reconstruction.py 96%, reconstruction_study.py 96%
• full suite: 265 passed, 98.26% total branch coverage
• native C++ matching-engine suite passed
• wheel and source distribution built; both passed twine check

Follow-up

The next increment can materialize each valid reconstruction segment to canonical Parquet plus a model manifest. A separate community validation issue should solicit paired L2/L3 evidence and alternative cancellation-allocation models.

Community benchmark: #62