[witbrock]

Summary

• surface workflow capability/index unavailability as a user-visible and replay-visible blocker
• prevent zero-candidate operational discovery failures from being cached as authoritative empty discovery
• preserve contract-named launchable workflow candidates when expected-outcome contracts list internal/low-level tools
• improve authenticated replay preflights and diagnostics around DB, LLM, auth, workflow capability, and path evidence

Root Cause

The bad arXiv replay route was not primarily a selector LLM mistake. The selector was handed a bad effective candidate set after workflow discovery produced a zero-candidate result while the capability index/query surface was transiently unavailable; that operationally invalid empty result could then be memoised/reused. A second support-layer issue let expected-outcome required_tools filter away a launchable workflow explicitly named by the represented contract.

Validation

• PYTHONPATH=tests/backend:. .venv/bin/python -m pytest tests/backend/test_orchestrator_selector_core_routing.py tests/backend/test_workflow_discovery_memo_service.py -q -> 32 passed
• .venv/bin/python -m pytest tests/test_run_authenticated_browser_workflow_replay.py tests/backend/test_mongo_dns_fallback_recovery.py -q -> 17 passed
• .venv/bin/python -m py_compile src/backend/integrations/internal_mcp/orchestrator.py src/backend/services/workflow_discovery_memo_service.py -> passed
• git diff --check -> passed
• exact authenticated arXiv replay selected and completed #V#arxiv_paper_representation_workflow, then delegated/read back/verified the represented paper and file-copy evidence

Notes

Stack order: merge this before jvnautosci-2561-replay-gmail-blocker-oracle. The 2561 branch fixes a replay oracle false positive discovered while validating this branch.