feat(product): ship transcript-first Braid terminal

[drewstone]

What ships

• Makes the selected AgentProfile authoritative for each agent-runtime turn, including runner, model, reasoning effort, and output limit.
• Restores trusted profiles after restart and switches the next run without changing the active receipt.
• Generalizes retained execution behind small lifecycle modules for CLI Bridge runs, analysis calls, reconnect, cancellation, and exact identity.
• Refines the terminal into a quiet transcript-first shell with profile selection, focused activity, cited /ask results, honest usage, and keyboard navigation.
• Adds real package, CLI Bridge, terminal, and Tangle Sandbox proof paths.

Direct proof

• pnpm check: 667 tests, 665 passed, 2 external-only rows skipped, 0 failed; 15/15 performance checks; 24/24 live-contract checks; CLI Bridge adversarial matrix passed.
• Exact packed Braid demo: Product engineer profile → Claude Code / Opus → workspace test task → Trace analyst profile / Sonnet → cited /ask; 3/3 findings supported.
• Tangle Sandbox production stress: 20/20 Braid → agent-runtime → OpenCode → GLM-5.2 runs at concurrency 4; latency min 24.671 s, median 27.981 s, p90 34.738 s, max 44.578 s; 20/20 cleanup; 0 leaked environments.
• 100,000-worker unchanged-screen p90: 0.95 ms under the full release run.

The exact cloud artifact is committed at artifacts/verification/live/tangle-sandbox-braid-execution-stress-production-20260812.json.

Honest boundary

The current Tangle cloud route is ephemeral. Braid does not claim crash-safe cloud reattachment, checkpoint migration, or cloud fork until the shared provider exposes exact recovery and control. That work remains tracked in agent-runtime #800 and agent-sdk #146. Missing normalized tool-result parts remain tracked in agent-runtime #762.