feat(tui): ship the agent runtime terminal

[drewstone]

Outcome

Braid is now a production terminal client over agent-runtime.
One portable AgentProfile controls the runner, model, reasoning effort, output limit, connection, and execution location.

The same application core powers the Pi TUI, plain terminal output, and JSONL control mode.
It adds durable conversations and forks, approvals, reconnect and cancellation, trace analysis, comparisons, worker activity, graphs, and separate direct, analysis, and worker accounting.

Real product proof

This is the packed package running through Local CLI Bridge to Pi 0.84.1 and GLM-5.2.
Pi edits a real workspace and passes 17 of 17 external tests.
/ask then returns three cited findings from six visible model calls across two pages.

• Repository: 610 passed, 0 failed, 2 protected checks skipped
• Performance: 15 passed, 0 failed
• Live demo: 18 passed, 0 failed
• Visual proof: 75 artifacts, 11 states, 4 terminal sizes
• Architecture: 454 modules, 2,323 edges, 0 cycles
• Independent Luna Max review: APPROVE, no remaining defects
• Secret scan: no leaks in 1,432,769 artifact bytes
• Exact tarball SHA-256: 32dda757d252bff4e120a8c69d776cb76b3ee971e3f2649d661875e77a20f7b8

Honest limits

Protected Tangle inference and sandbox checks require their real credentials and did not run locally.
agent-runtime issue 762 owns incremental normalized tool events.
agent-runtime issue 773 owns exact native execution and worker attachment.
Braid shows these limits and does not simulate support.