[witbrock]

Summary

• narrows the replay harness Gmail OAuth/profile blocker oracle so it does not scan all task evidence as one text blob
• ignores non-evidence LLM input/request surfaces such as llm_request, context_messages, and prompt_preview
• avoids treating identifier fields such as contract_id, fact_id, and workflow_id as Gmail failure context
• adds durable AGENTS guidance requiring path analysis for replay pass, fail, timeout, and typed-blocker outcomes

Root Cause

The 2560 validation replay completed correctly, but the harness classified the artifact as blocked because unrelated evidence leaves were combined: gmail_arxiv_* identifiers supplied gmail, while unrelated workflow/profile or LLM prompt text supplied auth/profile/token language.

Validation

• .venv/bin/python -m pytest tests/test_run_authenticated_browser_workflow_replay.py -q -> 18 passed
• .venv/bin/python -m py_compile scripts/run_authenticated_browser_workflow_replay.py -> passed
• git diff --check -> passed
• reclassified the real 2560 artifact: old blocker gmail_oauth_or_profile_blocker; new verdict pass, blocker null

Stack

This is stacked on PR #290 / jvnautosci-2560-capability-index-red-flag and should merge after 2560.