[codex] S3 cross-topic scheduled injection benchmark

[Joacocade]

Linked issue

S3 follow-up to S2 Issue 19 / PR #25. This PR is intentionally stacked on codex/s2-issue19-baseline so the diff only contains the S3 cross-topic extension.

Summary

Adds a reproducible S3 cross-topic scheduled-injection benchmark under backtesting/s3-cross-topic-injection/.

What changed:

• Added compact topic packets for football, Bolivia, and IPC.
• Added a 3 topics x 2 models x 7 conditions matrix.
• Added a resumable runner with prepared-simulation reuse.
• Added package validation, technical summarization, and deterministic metric extraction scripts.
• Added committed smoke/full summaries, condition metrics, run ledger, final report, and detailed result analysis.

Experimental design

Models:

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

Topics:

• football: Argentina vs Colombia, Copa America 2024 final
• bolivia: Bolivia 2025 presidential runoff
• ipc: Argentina IPC 2025 forecast

Conditions per topic/model:

• baseline-control
• signal-early
• signal-mid
• signal-late
• counter-signal-mid
• noise-near-mid
• noise-off-mid

Full matrix size: 42 runs.

Results

Technical validity:

• 42/42 full matrix rows valid.
• 6/6 baselines fired zero scheduled events.
• 36/36 injected conditions fired exactly one scheduled event.
• 6/6 topic/model pairs reused one prepared simulation across their seven conditions.

Directional findings:

• Football mostly reproduces the S2/V2 pattern: Gemma remains sticky toward Argentina, while Llama flips weakly to Colombia under counter-signal-mid.
• Bolivia is the cleanest cross-topic signal result: both models move toward Paz under signal injections and toward Quiroga under counter-signal.
• IPC remains stable toward lower/disinflation except under counter-signal, where Gemma becomes unclear and Llama flips to higher/rebound.
• Off-topic noise is generally ignored; near-topic noise is more dangerous, especially in Bolivia.

Conclusion

S3 supports the S2/V2 conclusion that scheduled injection is technically robust and directionally meaningful, but extends it across domains. The effect transfers to Bolivia and IPC, with topic/model-specific sensitivity. The strongest technical result is the reliable event audit; the strongest behavioral result is that relevant counter-evidence can change direction, especially for Llama.

This PR should be read as a technical and deterministic directional benchmark. Artifact-only ReportAgent scoring is documented as optional follow-up, not included in this PR.

Evidence

Primary committed evidence:

• backtesting/s3-cross-topic-injection/evaluation/final_s3_report.md
• backtesting/s3-cross-topic-injection/evaluation/results_analysis.md
• backtesting/s3-cross-topic-injection/evaluation/full_summary.md
• backtesting/s3-cross-topic-injection/evaluation/condition_summary_metrics.md
• backtesting/s3-cross-topic-injection/RUN_LEDGER.csv

Local reproducibility artifacts remain under runs/s3_cross_topic/* and are intentionally not committed because they include SQLite/log outputs.

How to test

Run from backend/:

uv run --frozen python -m py_compile ../backtesting/s3-cross-topic-injection/scripts/validate_s3_package.py ../backtesting/s3-cross-topic-injection/scripts/run_s3_matrix.py ../backtesting/s3-cross-topic-injection/scripts/summarize_s3_smoke.py ../backtesting/s3-cross-topic-injection/scripts/extract_s3_metrics.py
uv run --frozen python ../backtesting/s3-cross-topic-injection/scripts/validate_s3_package.py

Validated locally:

• script py_compile passed;
• package validation passed with topics=3 models=2 conditions=7 smoke_rows=12 full_rows=42.