feat(nlp): replace AnthropicEventExtractor with RuleBasedEventExtractor

[cipher813]

Summary

Applies the standing rule per `[[preference_llm_calls_confined_to_research_module]]` — LLM calls live in research; data uses existing metadata + rule-based classifiers. The news pipeline's Haiku-backed event extractor is removed and replaced with a deterministic classifier using two zero-cost signals already on the wire:

1. Vendor tags (`NewsArticle.tags`) — Polygon emits keywords, GDELT emits structured event codes, Benzinga emits Channels. We were paying Haiku to re-derive what Polygon / GDELT already tagged.
2. Title-keyword regex — backstop for sources that don't populate tags (Yahoo RSS). 17 pattern → category mappings against the closed `DEFAULT_EVENT_CATEGORIES` taxonomy.

Why this is the right answer, not a kill-switch

Code audit found the Haiku per-article structured output was aggregated to 5 scalar/list columns (`event_count`, `event_severity_max/mean`, `event_categories`, `top_event_descriptions`) before any research consumer touched it. The "zero-shot novel-event detection" capability was mostly wasted — research only sees per-ticker rollups. Tag-based + keyword-based classification produces equivalent rollups deterministically.

Cost impact

Retires the largest previously-untracked LLM cost slice per the Phase 0 audit estimate ($20–60/mo → $0). The research consumer sees identical EventFlag shape (extractor slug changes from `anthropic_haiku` to `rule_based`) and identical aggregate columns in `news_aggregates/{date}.parquet`.

Substrate cleanup

Retires three files added earlier this session — all dead code now that the LLM call site is gone:

• `collectors/nlp/event_extraction.py` — the Anthropic extractor itself
• `rag/pipelines/_cost_telemetry.py` — Phase 0.2 buffer (PR feat(cost-telemetry): wire news event extraction + lib v0.32→v0.33 (Phase 0.2) #308) + Phase 4 feat: data quality gates — parquet validation + email fixes #1 runaway-cost breaker (PR feat(cost-telemetry): runaway-cost circuit breaker on news pipeline (Phase 4 #1) #309)
• `tests/test_news_cost_telemetry.py` — mirrored tests

`DEFAULT_EVENT_CATEGORIES` moves into the new `rule_based_event_extraction.py` so the closed taxonomy stays accessible.

Protocol contract

`EventExtractor.extract` gains an optional `article_tags: tuple[str, ...] = ()` kwarg (back-compat default). Pipeline plumbs the tag union across article variants. Any future EventExtractor (FinBERT, spaCy, reactivated LLM via research module) consumes the same shape.

Severity convention

Rule-based flags emit `severity=0.5` uniformly (the EventFlag protocol's documented default). Haiku severity was a free-floating judgment never tuned by any operator alert. Per-category severity can be added via YAML if a downstream surface needs it.

Test plan

• `TestRuleBasedEventExtractor` (10 tests) — empty-text short-circuit, no-match, per-category title classification (earnings / M&A / FDA), tag-based classification, tag+title union, multi-category, deterministic ordering, zero-LLM contract, description shape
• Suite 1493 → 1479 net (retired 9 cost-telemetry + 7 Anthropic extractor tests; added 10 rule-based tests)
• Sat 5/30 SF: RAGIngestion fires, `news_aggregates/{date}.parquet` shows non-empty event_categories from rule-based path, `event_count` > 0 for tickers with vendor tags
• Research's substrate snapshot reads the new shape cleanly (no consumer-side changes required)

Composes with

• `[[preference_llm_calls_confined_to_research_module]]` — the rule this PR enforces
• alpha-engine perf(daily_append): chunked universe pass with gc.collect between chunks #212 (executor EOD narrative kill switch) — sibling application of the same rule. Two non-research LLM call sites: this PR retires data's entirely; executor's keeps the kill-switch substrate (default off) since the LLM path may be operator-reactivated
• Retires the substrate from data feat(cost-telemetry): wire news event extraction + lib v0.32→v0.33 (Phase 0.2) #308 + data feat(cost-telemetry): runaway-cost circuit breaker on news pipeline (Phase 4 #1) #309 — both became dead code with the LLM call site they instrumented

🤖 Generated with Claude Code