2026-03: Constrained autonomy, reality-aware replay, and shadow rollback

The investment stack now does more than expand themes and candidates.

It also constrains itself.

What changed

cross-source contradiction and rumor penalties now reduce confidence
stale mapping priors now decay instead of dominating current ideas
recent-evidence floors can stop live deployment
spread, slippage, liquidity, and session-state penalties now affect replay summaries
idea cards now resolve to deploy, shadow, watch, or abstain
weak recent tracked performance can arm shadow rollback

Where it shows up

Investment Workflow
Backtest Lab
docs/investment-usage-playbook.md
docs/automation-runbook.md

Why this matters

The system is still not an unconstrained auto-trader.

It is now closer to a constrained autonomous research and decision stack:

more automated than before
more explicit about when it should stand down
less likely to over-trust stale or low-quality evidence