Wheel expanded sample — fixed rules, more history

Run from the repository root with Python 3.11+ and numpy, pandas, psycopg[binary]:
  python -m research.wheel_expansion --output /absolute/new/folder
  python -m unittest tests.test_wheel_backtest tests.test_wheel_expansion

The default PostgreSQL DSN is postgresql://localhost/duke535_research.
The load_database function in homework3/wheel_backtest.py records the exact
OptionMetrics table fields and filters. Licensed source observations are not
redistributed. Input SHA-256 and file hashes are saved for each annual cell.

The current plan covers 22 stock-years across AMD, ADBE and AMZN. This scope was
narrowed after reviewing the earlier four-stock results; it is not a new blind
test. The prior full panel is archived locally. Retained books are unchanged.
Original and G012 rules stay fixed. Each year starts with $50,000 and ends flat.
The first usable stock session must have a prior stock session. Missing option
quotes prevent entry; missing held marks stop that account. Splits fail closed.
No-trade years are kept. Annual accounts are not one ten-year compounded fund.
No new optimization, dividend model, minute data, or observed fills are claimed.

Count original and G012 completed options separately. Option opens and exits
are not extra independent bets. The 72 old parameter trials are not new samples.
These are retrospective tests on three technology/growth names: correlated dates,
survivorship/universe selection and prior research overlap remain limitations.

Base assignment: >= $0.01 ITM at expiry; remove option and exchange 100 shares
at the strike in one atomic event. Optional assignment scenarios are NOT history:
1. Early assignment on an ITM held option with midpoint minus intrinsic <= $0.05.
2. Assign every expiry within 0.5% of strike, including slightly OTM expiries.
3. Do not assign those near-strike expiries, including slightly ITM expiries.
They measure sensitivity to alternative settlements, not calibrated probabilities
or guaranteed worst/best bounds. Current-close observations drive hypothetical
settlement events; strategy selection still uses prior completed sessions.
Actual American exercise can happen outside these flags. Dividend-paying stocks
need an explicit dividend calendar and account cash flows before inclusion.

Each scenario retains a separate full blotter, daily ledger, decisions and
monthly file in the bundle. Event fields identify hypothetical settlements.
The public account.json contains original and G012 books plus scenario metrics.
Prior published AMZN 2023–2024 results remain unchanged and separately labeled.
