feat(autoresearch): live trade bridge + all-symbol research (HCLI port 2/5)

[JaeLeex]

Part 2 of 5 of the HCLI → agent-cli port. Lets a user research → create → trade a strategy entirely from agent-cli, for crypto and BTCSWP + commodities. This is the foundation the House fleet (PR 3/5) builds on.

What's new

• hl strategy load <name> — runs an autoresearch strategy LIVE. AutoBarStrategy bridges the hourly on_bar(bar_data, portfolio) -> [Signal] contract to the tick engine: folds 10s ticks into hourly bars, calls on_bar once per boundary (wall-clock OR tick-count fallback for replay/mock), holds target between bars, and translates each Signal.target_position (signed USD) into delta = target − context.position_notional → |delta|/mid IOC order. Faithful port of ACC server/src/strategy-deployer.ts.
• hl strategy new|list|show|path|prepare — scaffold lifecycle.
• hl autoresearch run <name> [--iterations] [--agent demo|llm] — the Karpathy loop (scaffold → baseline → propose → eval → keep/discard, results.tsv + JSONL event log) + status|results|tail. Ports the nunchi-cli worktree command + scaffold template.
• cli/autotrader_bridge.py — Python port of ACC autotrader-bridge.ts results.tsv parser (best score/commit/#experiments).

All-symbol support (real coin names, discovered live — no fabrication)

The scaffold prepare.py gains an HIP-3 data adapter. Coin names were discovered from the live HL info API (perpDexs + meta/candleSnapshot), not guessed:

• BTCSWP → yex:BTCSWP (testnet-only; mainnet returns null — routed to the testnet endpoint).
• Commodities (mainnet xyz dex, confirmed returning candles): xyz:GOLD, xyz:SILVER, xyz:COPPER, xyz:PLATINUM, xyz:PALLADIUM, xyz:NATGAS, xyz:BRENTOIL, xyz:CL (WTI). Listed-but-no-data yet (WHEAT/CORN/URANIUM/ALUMINIUM) documented as add-once-liquid.
• Crypto (BTC/ETH/SOL) stays on CryptoCompare. Date splits made env-overridable (HIP-3 history is newer).

Guardrails (the live loop previously lacked these)

• Drift detection — rolling live realized return/Sharpe vs the kept backtest score; warns on divergence.
• Retrain trigger — one-shot retrain_callback; the runner writes a retrain.flag marker.
• reduce_only / safe_mode — suppress new/added risk (reducing orders still allowed).

Verification (real, measured)

• Imports clean; hl strategy + hl autoresearch + subcommands render.
• autoresearch run smoke --iterations 1 --agent demo ran the full loop and wrote results.tsv.
• strategy load smoke --mock --max-ticks 50 → bridge produced 15 orders, first fill = exactly 10% of $100k equity (sizing correct), 0 sent on held ticks.
• Real downloads + backtests: GOLD 3923 bars, BTCSWP 2672 bars; crypto sanity backtest score 2.57. (Naive momentum scores poorly on commodities — that's what the research loop is for; the scoring pipeline is structurally sound for HIP-3 symbols.)

Notes

• Based off main. Scaffold default keeps ACTIVE_SYMBOLS=[BTC,ETH,SOL]; commodities/BTCSWP are opt-in (data+adapter support all).
• Builder fee auto-attaches on live orders via the unchanged place_order path (same as PR 1/5).