Shadow idea-generation sidecar (first milestone of #43)

[OnePunchMonk]

First milestone of #43: a non-learning shadow sidecar. No RL, no new dependency, off by default (AGENTQUANT_IDEA_SIDECAR=1 to enable). The zero-key demo never imports it.

It replays a fixed heuristic policy or a seeded random policy over the canonical momentum grid and logs a versioned observation / action / reward trajectory, including propensities and every rejected candidate.

What is in src/agent/idea_sidecar.py:

• Typed, versioned schema: Observation, Action, Budget, RewardSpec, RewardComponents (components reported separately, composite is None when the holdout outcome is missing).
• Observations use dev-window data only, and failure records are kept only if timestamped on or before the episode's dev_end. Unparseable timestamps are dropped.
• Actions must be a canonical grid point from the step's candidate set, otherwise ActionOutOfSpace. The sidecar never emits code.
• Winner is picked by dev Sharpe with the same 0.3 gate as the search arms, and the sealed holdout is graded once after the trajectory closes. final_holdout=True goes through FinalHoldoutGuard.
• Every record carries an evidence tier (fixture, measured_historical, unverified; default unverified), a sidecar hash, and a replay check.

Tests (tests/test_idea_sidecar.py, 10 fixture tests) cover the five items from the issue: no future memory or holdout in observations, actions inside the canonical space, equal budget for random vs heuristic, replay reproducing logged decisions, and no bypass of the quality gate or holdout guard. Full suite passes locally (240).

Not in this PR, on purpose:

• Any learned policy or offline bandit/RL.
• The LLM-backed source (enum value is reserved, the candidate builder does not emit it).
• Wiring into RunManifest / research memo. Records are standalone JSON for now.
• The matched-budget baseline suite against grid, frozen agent and diagnosis mutation, and the paired uncertainty report.

All results here are fixture-tier on synthetic data. Nothing claims trading performance.