feat(leaderboard): add 14 external strategies as baseline classes

[TrentonNewWorld]

Summary

• Adds 14 deterministic (no LLM call) baseline strategies to the leaderboard registry, translated from three external sources backtested in a "Strategy Lab" report: TradingAgents' technical-composite proxy, 6 of QuantConnect's public Investment Strategy Library entries, and 8 of freqtrade-strategies' most prominent strategies (crypto-specific filters identified and stripped rather than blindly ported; GodStra excluded as not portable).
• Adds _signal_engine.py, which bridges these daily-scale signals (RSI-14, SMA-200, monthly rebalances, etc.) onto the hourly equity-curve contract BaselineStrategy.run() must honor — resamples hourly bars to daily OHLCV, evaluates each strategy's weight function once per trading day using only history strictly before that day (no look-ahead), and marks equity every hour in between.
• Every strategy's lookback adaptively caps at whatever history is actually available rather than requiring a fixed window, since the real leaderboard's contest window (~1 month) is much shorter than the full year these strategies were originally validated against.
• Registers all 14 in registry.py and adds matching entries to leaderboard.json's strategies array (a registered class alone is inert — it only becomes visible/live once it also has a config entry, since that's the only call site for the registry).
• Also includes two earlier, unrelated fixes made in the same working session: the Windows venv-path + repo-root-relative DATABASE_PATH bugs (fix(backtests)), and a risk-gated Alpaca live-trading broker + a new Mission Control wallet/holdings page (feat(trading)).
• Along the way, fixed two pre-existing gaps caught while verifying this: mission_control.py was leaking raw exception text via "error": str(e) (violates test_error_detail_sanitization.py's CodeQL-motivated guard), and test_app_composition.py's route contract didn't yet include the two Mission Control routes.

Test plan

• pytest dashboard/backend/tests/domain/leaderboard/test_external_strategies.py -v — 58 new tests (registry identity, key resolution, required_symbols, run() smoke tests against synthetic hourly bars, including a short-history graceful-degradation case).
• Full backend suite (pytest dashboard/backend/tests/) — only the same 8 pre-existing, unrelated failures remain (iFinD, vnpy, optional-SDK, test_engine_move, test_canonical_consumers, test_router_move::test_no_circular_imports) plus one pre-existing CSRF test error; confirmed via a clean git stash baseline that none are caused by this branch.
• Verified end-to-end against real Alpaca market data on a local dev server: all 26 leaderboard entries (12 original + 14 new) return sane, finite equity curves via GET /api/v1/leaderboard?period=contest, and all 14 new strategies appear by name in the frontend's chart legend/picker on the Competition tab.

🤖 Generated with Claude Code

[vercel]

@TrentonNewWorld is attempting to deploy a commit to the allan-feng's projects Team on Vercel.

A member of the Team first needs to authorize it.

[Allan-Feng]

Thanks for putting this together. We really appreciate the work that went into the strategies.

A suggestion: rather than adding all of them to the leaderboard right away, it would be great to pick 2–3 of them and run them through the ATL backtest workflow first, and compare their results with the existing market indices/ buy-and-hold baselines.

After testing the strategies, Agent Supermarket would be a great place for them. That’s where we’d love community agents to land, so people can try them, clone them, and iterate.

Happy to talk through details anytime. Feel free to ping me (allanfeng) on our Discord: https://discord.gg/9HnQ6XDG98

[FlyM1ss]

Review pass done; fixes pushed as three commits on this branch (b154da8d, 32e2c14e, 998ef916). Full backend suite green locally (3421 passed), CodeQL alert #1282 (py/unused-import) closed.

Strategies (b154da8d) — behaviour: turn_of_month bought SPY on the last day of every contest (last date flagged as month-end) and a zero open started a phantom hold; capm_alpha_ranking hardcoded 0.5 weight (half the book in cash with one qualifier) and its OLS variance was over the full window while the covariance used the symbol's valid days; pct_change() padded gaps into 0% days (capm, volatility_effect); bandtastic could never trade (50-day floor vs ~44-day window); trendrider's SMA-200 went permanently NaN after one gap day; rsi NaN with no down moves, adx NaN on a flat stretch; build_price_cache dropped any symbol missing the first timestamp for the whole run. Structure: hlhb/trendrider/supertrend_triple precompute indicators once per symbol (causal → bit-identical, pinned by test_indicators.py); daily_history maps dates once not 5×; dead lot_size removed; the 12 required_symbols + 14 num_trades copies moved into base.py. Tests: the multi-symbol smoke tests were vacuous (fixture SYM0..29 vs DJIA-30 default → run() returned [] before touching anything) — they now pass symbols and assert curve length; added a reference-window engine test.

Live broker / Mission Control (32e2c14e) — /api/v1/mission-control/overview is now admin-gated (it exposes the real account and whether execution is armed); live snapshot errors no longer 500 the page; 30 s cache; page sends credentials + escapes output. alpaca_live_service runs the decision through the LLM validator before building orders; feed resolution errors are fatal instead of swallowed; ALPACA_LIVE_* stripped in conftest; alpaca_live.json.example + CLAUDE.md docs.

Backend (998ef916) — shared resolve_python_exe/resolve_env_path replace four copies of the Windows-venv / relative-DATABASE_PATH logic; leaderboard service builds _cached_run_index once per request (was _find_cached_run × 19 entries × 3 call sites) with a single-flight lock per window; leaderboard.js styles preset-less baseline entries as strategy, not team.

Not changed, flagged for follow-up

• Empty weights {} = liquidate (sanctioned by the _signal_engine docstring); entry/exit strategies keep held instead — worth a deliberate decision.
• trendrider's fixed ema(close, 50) is all-NaN below 50 rows, so the golden-cross entry / bearish-cross exit never fire on the leaderboard window (the other four signals still do). Same cap as bandtastic would fix it but breaks the precompute.
• TEAM_COLOR_PALETTE has 10 colours for 19+ entries → collisions; presets for the new labels would need to avoid MODEL_COLOR_PALETTE.
• Truthy-env parsing is duplicated across the two live services and the leaderboard flags — a shared helper.
• User-facing docs now stale (not edited here): docs/source/lab/live_trading.rst intro, docs/source/lab/getting_started.rst (~L130–154), README "Go live" bullet.

[FlyM1ss]

Overall looks good. I also recommend to test 2 to 3 strategies first.