Add news & filings sentiment brief skill to equity-research vertical

[HumphreySun98]

What

Adds a news-sentiment skill to the equity-research vertical: it turns a corpus of unstructured financial text — news headlines, press releases, and filing excerpts (8-K / 10-Q / 10-K) — into a structured, fully source-cited sentiment brief for a ticker or watchlist.

Each item gets a sentiment label from a fixed set (positive / neutral / negative / mixed), a verbatim source quote, entity/relation extraction, and an event tag; the skill then aggregates the distribution, logs notable events, and surfaces a human-review queue.

Why

equity-research covers earnings analysis, initiating coverage, thesis tracking, and morning notes — but nothing systematically labels a corpus of news/filings item-by-item with per-item sentiment and citations. earnings-analysis is a single-quarter deep report; morning-note and thesis-tracker consume conclusions. This is a clean, non-overlapping gap, and it plays to Claude's core strength: reading unstructured text and producing auditable structure.

The design is inspired by the financial-NLP task decomposition popularized by open-source finance LLM work (sentiment scoring / entity-relation extraction / event tagging) — recast here as inference-time instructions, not model training, and kept strictly inside the repo's guardrails: no price target, no forecast, no buy/sell call; every label is quote-backed and staged for human sign-off.

What's inside (mirrors the dcf-model gold-standard file pattern)

• SKILL.md — MCP-first data sourcing; a "no forecast / no recommendation" boundary; the requirement that every label carry a verbatim quote; the labeling workflow; the output schema; correct patterns / common mistakes; and guardrails.
• references/labeling-guide.md — the label set with a single shared score boundary (±0.15, disjoint and exhaustive), the confidence scale, the entity/relation schema, the event taxonomy, edge cases (negation, forward-looking vs realized, company vs analyst voice, filing boilerplate, headline/body divergence, hindsight leakage), worked examples, and the aggregation rule.
• scripts/validate_sentiment.py — an openpyxl validator (mirrors the dcf-model validator) that checks the required columns, the sentiment enum, a non-empty source_quote on every item, sentiment_score in [-1, 1] (error) and consistent with its label (warning), and — crucially — that low-confidence and mixed items are flagged needs_review = yes.
• TROUBLESHOOTING.md and requirements.txt (openpyxl only).

Validation

• python3 scripts/check.py → OK — 80 file(s) checked, 0 issues.
• scripts/validate_sentiment.py tested against clean and deliberately broken workbooks — the broken cases correctly flag an invalid label, an empty source quote, an out-of-range and label-inconsistent score, an over-magnitude mixed score, and unflagged low-confidence/mixed items.
• The pre-commit hook auto-bumped equity-research plugin.json 0.1.2 → 0.1.3. The new skill is not bundled into any agent, so there is no sync drift.

Frontmatter is name + description only, matching the existing vertical skills. This PR is self-contained and touches only the new skill directory plus the automatic version bump.