feat: sentence-level granularity and weighted multi-signal scoring

[Jamie-BitFlight]

Summary

Two related architectural improvements:

1. Sentence-level granularity — score each sentence independently instead of whole-response, enabling precise flagging ("sentence 3 of 7 has high hallucination risk")
2. Weighted multi-signal aggregation — combine scores from multiple detection categories into a single weighted score per sentence

Current Behavior

• Detection operates at whole-response level
• Each trigger match is independent — no aggregation
• No sentence-level breakdown

Proposed Behavior

Sentence Splitting

// Regex-based sentence splitter (zero dependencies)
const sentences = text.split(/(?<=[.!?])\s+/);

Per-Sentence Scoring

Each sentence gets scored by all active detection categories:

{
  sentence: "I think the issue is caused by a race condition.",
  index: 2,
  scores: {
    speculation_language: 0.8,  // "I think"
    causality_language: 0.7,    // "caused by" without evidence
    pseudo_quantification: 0.0,
    completeness_claim: 0.0,
  },
  aggregateScore: 0.68,  // weighted average
  label: "UNCERTAIN",     // GROUNDED < 0.30 < UNCERTAIN < 0.60 < HALLUCINATED
}

Configurable Weights

// Default weights (configurable via .hallucination-detectorrc.cjs)
const DEFAULT_WEIGHTS = {
  speculation_language: 0.25,
  causality_language: 0.30,
  pseudo_quantification: 0.15,
  completeness_claim: 0.20,
  fabricated_source: 0.10,  // if #18 implemented
};

Three-Tier Labels

Label	Score Range	Meaning
GROUNDED	< 0.30	Sentence is evidence-based
UNCERTAIN	0.30 - 0.60	Mixed signals — may need verification
HALLUCINATED	> 0.60	High confidence of hallucination

Acceptance Criteria

• Sentence splitting works for standard English text
• Per-sentence scoring for all active categories
• Weighted aggregation with configurable weights
• Three-tier labeling (GROUNDED / UNCERTAIN / HALLUCINATED)
• Backward-compatible — existing whole-response mode still works
• Tests cover sentence splitting edge cases and aggregation math

Related Issues

• feat: confidence scoring (0-100) per trigger match #12 (confidence scoring — this subsumes it with richer architecture)
• feat: config file handling — cascading settings from multiple sources #10 (config — weights and thresholds configurable)

References

• ObvioSpectre/hallucination-detector scoring/aggregator.py (weighted aggregation)
• ObvioSpectre/hallucination-detector core/sentence_splitter.py (regex splitter)

[Jamie-BitFlight]

Grouping: Group D — Scoring Architecture

Implement together with: #12

Rationale: This issue subsumes #12 (as noted in the issue body). The weighted multi-signal scoring naturally produces a per-match confidence value. Implementing both as one unit is the correct approach.

Dependencies:

• Benefits from feat: config file handling — cascading settings from multiple sources #10 (config) for configurable weights and thresholds
• Benefits from Group C (feat: add self-contradiction detection #4, feat: self-contradiction detection via negation polarity #19, feat: degenerate repetition / collapse detection #20) for the sentence-splitting infrastructure — or can build its own splitter that Group C later reuses
• The three-tier labeling (GROUNDED/UNCERTAIN/HALLUCINATED) provides the framework for feat: corrected text suggestions alongside flags #13 (suggested fixes) to key off