Add same-day TSLA intraday strategy (tsladay profile)

[utkarshp845]

New market profile alongside the existing swing config: config/paper_tsladay.env and config/live_tsladay.env (bot/profile.py market "tsladay"), wired into docker-compose.yml (trade-tsladay/paper-tsladay/paper-tsladay-monitor).

This is a deliberately different strategy from the spy-market swing config (~1 trade/5-6 weeks, multi-day holds), built because that config doesn't capture TSLA's frequent intraday swings by design:

• Bidirectional (ALLOW_SHORTS=true): a position that's always flat by close carries no overnight gap risk, so shorting TSLA's down-legs is as safe as going long its up-legs, and backtests about as profitable (short side actually outperformed: $16.29/30 trades vs long $8.58/25).
• Capped at MAX_TRADES_PER_DAY=1: the account is $150 in a cash account, so more than ~1 buy+sell cycle/day risks reusing unsettled (T+1) proceeds and tripping a good-faith violation. This is an account-mechanics constraint the strategy is designed around, not a signal-quality choice.
• 15m bars, 2h intraday regime gate (vs the swing config's daily-EMA gate), 09:45-15:30 ET entry window, FLATTEN_BEFORE_CLOSE_MINUTES=15.

Backtest: bot/optimize_strategy.py walk-forward grid search fed 15m TSLA bars via yfinance (no Alpaca keys available outside EC2; free intraday history caps at ~60 trading days, so this is a much thinner, single-regime validation than the swing config's 2.9yr one - flagged throughout). Shipped config (65%/70% target/max notional, trimmed down from what the 60-day backtest could have supported as a margin of safety): 55 trades, win rate 53%, PF 2.01 (1.99 under 2x slippage), net +$20.01 (+13.3%) on $150, max drawdown 3.7%, positive in 3 of 4 walk-forward test windows. Full methodology, sizing sweep, and caveats: docs/strategy_tsla_day_2026-08.md.

Also fixes a replay/live parity gap this surfaced: bot/research.py::run_replay never simulated ALLOW_OVERNIGHT_HOLDING=false / FLATTEN_BEFORE_CLOSE_MINUTES (only the live loop did, via bot/trade_controls.py::evaluate_session_exit) - every prior backtest had overnight holding enabled so it never mattered. run_replay now calls evaluate_session_exit every bar a position is open. Regression tests in tests/test_research_replay.py. Also made two optimizer knobs configurable that were previously hardcoded for swing-style signals (bot/optimize_strategy.py): ALLOW_SHORTS as a grid dimension (OPT_ALLOW_SHORTS_VALUES) and the trades/day acceptance band (OPT_SCORE_MIN/MAX_TRADES_PER_DAY, OPT_ACCEPT_MIN/MAX_TRADES_PER_DAY) - both default to prior behavior, no change for existing configs.

Recommendation: paper only (paper-tsladay) for a real stretch across different market conditions before ever considering live capital, per the doc.