fix(agents): preserve queued prompt across bounded context

[HomenShum]

Summary

• validate the exact queued user prompt before routing, reservation, or provider execution
• pass that validated prompt explicitly while keeping provider history bounded at zero
• fail closed on stale or malformed prompt rows
• redact SDK error-class wrappers and map internal empty-prompt failures to a safe retry message

Production evidence

The first bounded production dogfood after #529 selected the correct tier-eligible model and released its reservation, but the SDK received an empty provider message list because recentMessages was zero and only promptMessageId was supplied.

Verification

• focused FastAgent/runtime tests: 63/63
• full FastAgentPanel suite: 213 passed, 19 skipped
• app TypeScript: green
• Convex TypeScript: green
• production build: green
• git diff --check: green
• independent prompt-correctness review: approve
• independent accounting/lease/cancellation review: approve
• independent error-redaction review: approve

Release

P0 CI-gated auto-merge only. No admin bypass. No manual Convex production deploy.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
nodebench-ai	Ready	Preview, Comment	Jul 14, 2026 4:44pm

[github-actions]

Tier B regression failed against preview

Preview URL: (preview URL unresolved)
Run: 29350564417

Likely a kit-parity DOM regression. Check the Playwright report artifact attached to the run for the failing selector + screenshot.

This check would have caught all 4 A9 sessionRows iterations as a single PR — that's why it exists.

[github-actions]

Demo: walkthrough of the surfaces this PR changed is available as a workflow artifact (pr-demo-531) at https://github.com/HomenShum/NodeBenchAI/actions/runs/29351247848