feat(autopilot): scaffold signal engine + auto-link backtest metrics (Phase 3)

[Robin1987China]

Summary

Research Autopilot Phase 1 (#260) bridged hypothesis → goal and hypothesis → config.json, but left the rest of the loop wired by hand. Two manual steps remained the most error-prone parts of an autopilot run:

1. Writing code/signal_engine.py to the runner contract (a no-arg-constructible SignalEngine with generate(self, data_map)). Getting it wrong is the most common reason a backtest fails its pre-flight check.
2. Recording results — link_backtest needs the agent to open run_card.json and hand-extract the metrics dict.

This PR adds two deterministic, backend-only tools that close the loop.

Changes

• agent/src/tools/autopilot_tool.py:
  • scaffold_signal_engine — writes a contract-correct code/signal_engine.py stub with a runnable flat (no-position) default and the hypothesis signal_definition embedded as a docstring for the agent to implement against. Refuses to overwrite an existing engine unless overwrite=true.
  • link_autopilot_backtest — reads run_card.json from a run dir, extracts the scalar metrics, and links them to the hypothesis via the existing link_backtest, returning the metrics for thesis evaluation. Degrades to empty metrics with a warning if the key is absent.
  • Both resolve run_dir through safe_run_dir, inheriting the allowed-run-root boundary (fail-closed on path traversal, consistent with the backtest tool).
• agent/tests/test_autopilot_phase3.py: 12 tests.

Out of scope (deliberate)

• Compiling the free-text signal_definition into working signal logic (LLM territory; the stub is a starting point only).
• Auto-invoking the backtest tool — execution stays an explicit agent step so run cost and data-source side effects remain visible.
• Writing goal_evidence rows — the existing add_goal_evidence tool handles that and needs the per-turn expected_goal_id.

Test Plan

• ruff check agent/src/tools/autopilot_tool.py agent/tests/test_autopilot_phase3.py — clean
• pytest agent/tests/test_autopilot_phase3.py agent/tests/test_autopilot_tool.py -q — 19 passed
• Full suite excl. e2e — 3268 passed, 1 skipped, no regressions
• Stub-contract test imports the generated signal_engine.py and runs generate(), asserting no-arg construction + dict-of-Series output

Risk / rollback

Additive only; both tools auto-register via BaseTool.__subclasses__(). No schema, config, or behavior changes to existing tools. Revertable with a single git revert.

[warren618]

Merged — thanks @Robin1987China! 🙏 scaffold_signal_engine + link_autopilot_backtest close the two most error-prone manual steps of an autopilot run, and the generated stub matches the runner's SignalEngine contract exactly.