fix(runtime): make live provider outcomes truthful

[drewstone]

What changed

• Route Tangle sandbox execution through the current agent-runtime executor contract.
• Fail closed on raw sandbox errors, contradictory success markers, approval requests, questions, and plan decisions.
• Pass cancellation into sandbox creation and preserve usage, cost, cleanup, and environment observations.
• Store and hash one secret-safe materialization receipt without retaining the raw idempotency key.
• Use separate inference, sandbox, and analysis credentials in the release workflow.
• Make generated live credentials use Braid's production protected store with retryable cleanup.
• Report failed, unavailable, and uncaptured checks honestly.
• Lead the README with a real coding session through agent-runtime, CLI Bridge, OpenCode, and GLM-5.2.

Why

The previous sandbox adapter composed obsolete provider wrappers and trusted Runtime completion even when the buffered provider stream had failed or requested interaction.
That could report success after a failed turn or delete an ephemeral sandbox while presenting an interaction that could not resume.

Direct use

• The actual Braid TUI returned the exact requested response through agent-runtime → CLI Bridge → OpenCode → GLM-5.2.
• Run run-20518bf6-8092-4ec4-ba97-320e9e37b122 showed 16,109 input tokens, 24 output tokens, one model call, and 3,949 ms model latency in /activity.
• A packed headless run returned LIVE_BRAID_GLM-5_2_OK, exercised cancellation, and removed every temporary package, install, target, state, and credential path.
• Real Tangle sandbox execution created and read a workspace file, while a 5/5 concurrent stress run created unique sandboxes and confirmed all five were destroyed.
• CLI Bridge write-jail still blocks instruction-bearing OpenCode profiles; the upstream defect is tracked at fix(opencode): profile instructions fail before model call under write-jail drewstone/cli-bridge#145. The visible local run used an isolated loopback Bridge with that jail disabled.

Verification

• pnpm check
• 625/625 executed unit tests passed; 2 skipped.
• 15/15 performance tests passed.
• 2/2 native SQLite tests passed.
• 19/19 terminal and demo tests passed.
• 14/14 protected live-path tests passed.
• Live Bridge adversarial matrix passed.
• Independent focused review: no actionable findings.

Visual