Add agent nudge control

[BakedSoups]

Summary

Currently, once the wizard agent is running, the user can mostly observe or cancel. If the agent appears to be looping, over-investigating, or stuck on a dead end, there is no lightweight intervention control to steer it forward without ending the run.

This PR adds an n key binding on the run screens that nudges the active agent to finish the current line of work and move on if it is stuck. The nudge is streamed into the active Anthropic SDK prompt as a follow-up user message, so it works during an in-flight run instead of waiting for a restart.

This is similar in spirit to Claude Code-style human intervention, but deliberately uses an explicit nudge agent action instead of overloading Esc or Ctrl+C, which users reasonably expect to mean interrupt, cancel, or exit.

I made this a PR since its something I personally found useful

Changes

• add an n / nudge agent key binding on normal and audit run screens
• store nudges in the TUI store with monotonic ids and analytics
• stream nudges into the active Anthropic SDK prompt as follow-up user messages
• add store coverage for nudge sequencing, waiting, and abort cleanup

Test plan

• pnpm build
• pnpm vitest run src/ui/tui/__tests__/store.test.ts
• pnpm fix
• pnpm test

[gewenyu99]

Oh that's really cool. I had planned this here:
#769

[gewenyu99]

I need to do some rewiring for this to work with all the runners, but this is really really cool.

[BakedSoups]

Awesome, Thank you for checking out my pr.

[BakedSoups]

Just as a little extra thing to do,

I gave your idea of making nudges work across all runners a try by threading the TUI nudge source through the shared harness inputs.

Both linear and orchestrator flows now pass the nudge source to the selected runner. Anthropic continues using its existing prompt stream, while Pi forwards nudges through agentSession.steer().

Feel free to check it out when you have a chance.

verified with pnpm build, pnpm test, and pnpm fix.
29e1f53