- get_past_context returned every resolved lesson regardless of the run date, so
a historical run could learn from an outcome that had not happened yet
- record each resolved entry's resolution date (the last price bar used) and
filter get_past_context(as_of=trade_date) on it for a historical run; a
current-date run passes None so live behavior and pre-migration entries (no
stored resolution date, conservatively excluded from backtests) are unaffected #1251
