Derive runtime deadlines from Nitro maxDuration

[dcramer]

Objective

Make juniorNitro({ maxDuration }) the single source of truth for Junior's
serverless execution budget.

Simple contract

Junior should not require operators to configure the same timeout twice.

When an app sets:

juniorNitro({ maxDuration: 800 });

Junior should use that value both for the generated Vercel function
configuration and for runtime deadline calculations.

• The agent hard deadline leaves enough time to save state before Vercel stops
  the function.
• The conversation worker soft-yields earlier than the hard deadline.
• Queue visibility lasts longer than the function can run.
• AGENT_TURN_TIMEOUT_MS may make an agent run shorter, but never longer than
  the host budget.
• Production may raise maxDuration to the limit available to its Vercel
  project without separately changing runtime environment variables.

Scope

• Inject the build-time maxDuration value into Junior's generated runtime
  configuration.
• Derive the request deadline, conversation-worker soft yield, and queue
  visibility timeout from that value using explicit safety buffers.
• Remove FUNCTION_MAX_DURATION_SECONDS and the legacy
  QUEUE_CALLBACK_MAX_DURATION_SECONDS runtime configuration once the generated
  value is authoritative.
• Preserve a safe default when juniorNitro() does not receive maxDuration.
• Add configuration tests proving build and runtime values cannot drift.

Completion

• One maxDuration setting configures both Vercel and Junior runtime budgets.
• Agent execution stops before the host timeout and persists a continuation.
• Conversation work soft-yields before the hard deadline.
• Queue visibility remains longer than the configured function duration.
• A lower AGENT_TURN_TIMEOUT_MS override remains supported.
• Production can increase maxDuration without adding a matching environment variable.

Non-goals

• Automatically discovering Vercel account limits at runtime.
• Choosing the production duration value in this issue.
• Changing turn slice limits or retry policy.