Build queue-aware market-making flagship

[Fink692]

What changed

• centers Quant Systems Lab on a queue-aware market-making flagship using the verified public LOBSTER AAPL Level-5 sample
• adds event-level ingestion, provenance manifests, order-book reconstruction and reconciliation, chronological calibration, and a shared replay engine
• compares fixed-spread, Avellaneda–Stoikov, queue-aware, toxicity-aware, and latency-aware policies under common fill, fee, inventory, liquidation, and accounting assumptions
• publishes queue, fee, and latency sensitivity analysis plus explicit failure cases and honest negative held-out results
• strengthens the valuation-regime study with a later partial test fold, block-bootstrap intervals, deflated Sharpe, an overfitting diagnostic, parameter stability, bond-sleeve analysis, and risk-matched baselines
• adds immutable experiment configuration, append-only run provenance, model cards, assumption and completion audits
• adds an executed start-here notebook, Streamlit dashboard, architecture diagram, paper PDF, tear sheet, narrated five-minute demo, and reproducibility workflow
• upgrades CI with Python 3.11–3.13, Ruff, Black, MyPy, coverage enforcement, dependency auditing, pre-commit, docs validation, and benchmark regression

Why

The repository already had broad model coverage. This change turns the strongest market-microstructure scaffolding into a focused, auditable research project and separates genuine research evidence from presentation claims.

The public sample is treated as pipeline validation, not proof of persistent profitability. All five held-out base market-making policies lose money, and the valuation study's bootstrap alpha interval crosses zero while simpler risk-matched baselines remain competitive or superior.

Impact

The README now provides a guided hiring-manager path through one flagship project. Raw order-book data remains ignored and is reproducibly fetched with verified hashes. Published artifacts are versioned, while licensed multi-session L2/L3 and international point-in-time datasets remain explicit future empirical gates.

Validation

• 195 tests passed
• 88.54% coverage (85% floor)
• Ruff passed
• Black passed
• scoped MyPy passed
• strict MkDocs build passed
• all pre-commit hooks passed
• dependency consistency passed
• no known audited dependency vulnerabilities
• notebook executed without error outputs
• dashboard loaded through Streamlit AppTest
• paper and tear-sheet PDFs rendered and visually inspected
• narrated 1080p demo validated at 5:14 with audio