fix: independent temporal-leakage verifier for search_web

[ethancjackson]

Summary

search_web (the shared news/context-retrieval tool in agent_factory.py, used by all 5 forecasting domains) was leaking post-cutoff information into historical backtests despite an explicit natural-language cutoff instruction — confirmed directly in a Langfuse trace (a5d33fccd8c99929c3784f774dda34e6, cutoff 2026-06-01, result citing WTI/Brent prices "by mid-June 2026"). The existing ContextRetrievalConfig docstring already called this out as "soft (LLM-judgment-based)... not a hard guarantee." A prior attempt at a hard date-filtered search API (Tavily) also leaked, which matches published research: date-restricted search is insufficient because underlying page content/metadata gets updated after original publish dates.

Clickup Ticket(s): N/A

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📝 Documentation update
• 🔧 Refactoring (no functional changes)
• ⚡ Performance improvement
• 🧪 Test improvements
• 🔒 Security fix

Changes Made

• Added an independent temporal-leakage verifier to _build_search_tool in aieng-forecasting/aieng/forecasting/methods/agentic/agent_factory.py: after every grounded search_web call with a cutoff date, a separate LLM call (verifier_model, defaults to ADVANCED_MODEL — distinct from search_model — so it doesn't share the same knowledge-attribution blind spot) extracts discrete factual claims and judges each against the cutoff, strips violations, and reports a confidence score.
• Retry loop: on a non-clean/low-confidence verdict, search_web retries (same query, with the previously flagged claims injected as explicit negative feedback) up to verifier_max_attempts (default 3) before returning an explicit [SEARCH_VERIFICATION_FAILED] ... sentinel — never silently returning unverified content.
• New ContextRetrievalConfig fields: verifier_model, verifier_max_attempts (3), verifier_confidence_threshold (8/10 — kept tunable rather than hardcoded since LLM self-reported confidence isn't well-calibrated).
• implementations/energy_oil_forecasting/analyst_agent/agent.py: threaded the three verifier knobs through all four news-enabled builders; hardened the search sub-agent's own instruction (reason step-by-step about claim recency from content, not source metadata; never supplement from background knowledge) as cheap defense-in-depth; root analyst instruction now explicitly handles the [SEARCH_VERIFICATION_FAILED] sentinel by proceeding on price history alone instead of guessing.
• 8 new unit tests in test_agent_factory.py covering immediate-accept, retry-then-accept (with feedback-injection verified), exhaustion→sentinel, both skip conditions (no cutoff / enforce_cutoff=False), default-model selection, configurable threshold, and verifier parse-failure handling.
• Verifier calls get their own Langfuse span for free (same litellm.acompletion callback path as the existing search call) — no extra instrumentation needed.

Testing

• Tests pass locally (uv run pytest tests/)
• Type checking passes (uv run mypy <src_dir>)
• Linting passes (uv run ruff check src_dir/)
• Manual testing performed (describe below)

Manual testing details:

• uv run pytest aieng-forecasting/tests/aieng/forecasting/methods/agentic -q — 89 passed (including the 8 new leakage-verifier tests), no regressions.
• uv run ruff check clean on all three changed source files.
• Verified against live production traces post-fix: confirmed the verifier now runs as its own span on every search_web call for both gemini-3.1-flash-lite-preview and gemini-3.5-flash news-agent variants (previously 1 child LLM call per search_web, now 2: search + verify). Directly observed it flagging and stripping borderline post-cutoff claims (e.g. "Brent trading in a range of approximately $65–$70 per barrel in September 2025" against a 2025-09-01 cutoff) and accepting the filtered result with confidence: 9. Sampled 200 search_web calls from the last 3 days across both models: 0 occurrences of the [SEARCH_VERIFICATION_FAILED] sentinel, no latency evidence of retries in that sample (so the retry-exhaustion path is implemented and tested but not yet observed in production).

Related Issues

N/A — follow-up to #160, which merged before this leakage issue was caught.

Deployment Notes

NB04 (04_systematic_backtest_eval.ipynb) is committed here mid-flight: it's actively re-running with the guard in place after clearing only the two news-agent predictors' cached results (all other predictors, e.g. LightGBM, stayed cached and untouched). Only the completed gemini-3.1-flash-lite-preview 2025-backtest prediction file is included in this commit; the gemini-3.5-flash backtest and both models' 2026-eval prediction files will follow in a follow-up commit once that run finishes — hence this is opened as a draft PR.

Checklist

• Code follows the project's style guidelines
• Self-review of code completed
• Documentation updated (if applicable)
• No sensitive information (API keys, credentials) exposed