[codex] S2 Issue 19 scheduled injection and ReportAgent evaluation

[Joacocade]

Linked issue

Closes #19

Summary

• Add scheduled intra-run Reddit injection support for Issue 19 using event_config.scheduled_events.
• Add the S2 Issue 19 benchmark packet, acceptance response, evidence policy, V1/V2 evaluation artifacts, and DeepInfra robustness artifacts.
• Add ReportAgent artifact_only mode so per-condition reports can be generated from isolated evidence bundles without shared graph/Zep tool reads.
• Generate artifact-only ReportAgent outputs for Qwen, Gemma, and Llama across all six V2 conditions.

Evidence

• Issue checklist: backtesting/case-a-s2-positional-noise/ISSUE_RESPONSE.md
• V1 report: backtesting/case-a-s2-positional-noise/evaluation/final_issue_report.md
• V2 report: backtesting/case-a-s2-positional-noise-v2/evaluation/final_v2_report.md
• DeepInfra report: backtesting/case-a-s2-positional-noise-v2/evaluation_deepinfra/final_deepinfra_report.md
• ReportAgent manifest: backtesting/case-a-s2-positional-noise-v2/evaluation_report_agent/report_agent_manifest.csv

ReportAgent manifest status: 18 rows, 18 completed, 0 errors.

How to test

From backend/:

uv run --frozen pytest ../tests/test_report_agent_resilience.py -q

Expected result: 23 passed.

ReportAgent artifact-only verification already completed for:

• qwen/qwen3-8b
• google/gemma-3-27b-it
• meta-llama/Llama-3.3-70B-Instruct-Turbo

The committed manifest at backtesting/case-a-s2-positional-noise-v2/evaluation_report_agent/report_agent_manifest.csv has 18 rows, 18 completed, and 0 errors.

Also verified no <tool_call>, <tool_code>, Thought, or tool-failure leakage in committed ReportAgent outputs.

Notes

The runs/ SQLite/log artifacts remain local reproducibility evidence and are intentionally not committed. The committed PR evidence is the compact summaries, metrics, narrative scores, final reports, and ReportAgent outputs.