fix(agentic): report exhausted responses clearly

[OllieinCanada]

Summary

Prevent exhausted WTI analyst runs from surfacing as misleading JSON parse
failures, and give the code-execution preset enough response headroom to finish
tool work and return its structured forecast.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• 📝 Documentation update
• 🧪 Test improvements

Changes Made

• raise a descriptive error when AgentPredictor receives an empty response
• increase the WTI code-execution preset default from 16K to 32K output tokens
• document the behavior and keep the concierge source snapshot synchronized

Testing

• Relevant tests pass locally (16 passed)
• Type checking passes (local full-package run was interrupted after an excessive runtime; CI will verify)
• Linting passes (ruff check on all changed Python files)
• Manual testing performed (describe below)

Manual testing details:

Both regressions were first run against the original implementation and failed:
the empty response leaked JSONDecodeError, and the WTI preset returned 16,384.
The same focused tests pass after the fix. No model, API, or external service was
used.

Related Issues

Fixes #180

Deployment Notes

No migration or deployment change is required.

Checklist

• Code follows the project's style guidelines
• Self-review of code completed
• Documentation updated
• No sensitive information exposed