feat: Intelligence Pipeline v3 — full implementation

Multi-stage evidence-grounded inference architecture replacing the
monolithic 9B model extraction pipeline. CPU-first specialist services
handle routine extraction while the 9B vLLM model is preserved for
semantic adjudication of ambiguous cases.

Key components:
- Capability-aware inference gateway (OpenAI-compatible + Ollama)
- Endpoint registry with DB migrations and REST API
- Sentence-aware document segmenter (property tests)
- Deterministic financial parsing with offset integrity
- Symbol resolution with ambiguity detection
- Specialist service (GLiNER2, dynamic batching, K8s deployment)
- Company-specific sentiment (FinBERT, calibration)
- Retrieval-based novelty and duplicate detection
- Confidence calibration pipeline
- Deterministic routing engine (property tests)
- 9B adjudication layer with VRAM gating
- Stock-specific impact model (features, labels, baseline, trained)
- Pipeline orchestrator (state machine, queues, leases, feature flags)
- Bounded parallelism (async workers, semaphore, load shedding)
- Observability (tracing, metrics, alerts)
- Compatibility adapter (v3→v2 golden mapping tests)
- Shadow/canary promotion framework
- Active learning and fine-tuning pipeline

Test results: 1,161 tests pass, ruff lint clean.
All 282 spec tasks completed.

services/intelligence_pipeline_v3/sentiment/aggregation.py

@@ -0,0 +1,136 @@
+"""Sentiment aggregation logic.
+
+Aggregates per-text sentiment scores into a single company result,
+with mixed sentiment detection when evidence groups disagree.
+
+Mixed sentiment is NOT an unconstrained fourth softmax label — it is
+computed from disagreement between evidence texts (some positive,
+some negative with margin > threshold).
+"""
+
+from __future__ import annotations
+
+from services.intelligence_pipeline_v3.sentiment.models import (
+    CompanySentimentResult,
+    TextSentiment,
+)
+
+# Disagreement threshold: both positive and negative probabilities
+# must be >= this value across evidence texts to signal disagreement.
+# A text with positive_prob >= threshold AND another text with
+# negative_prob >= threshold indicates conflicting evidence.
+MIXED_DISAGREEMENT_THRESHOLD = 0.3
+
+
+def aggregate_evidence_sentiments(
+    company_id: str,
+    per_text_scores: list[TextSentiment],
+    model_version: str,
+    calibration_version: str = "uncalibrated",
+) -> CompanySentimentResult:
+    """Aggregate per-text sentiment scores into a company-level result.
+
+    Computes weighted average probabilities and detects mixed sentiment
+    from evidence-group disagreement. Mixed is triggered when at least
+    one text has positive_prob >= threshold AND at least one other text
+    has negative_prob >= threshold.
+
+    Parameters
+    ----------
+    company_id
+        Resolved company identifier.
+    per_text_scores
+        List of TextSentiment objects, one per evidence text.
+    model_version
+        Sentiment model version string for lineage.
+    calibration_version
+        Calibration artifact version.
+
+    Returns
+    -------
+    CompanySentimentResult
+        Aggregated result with label, probabilities, per-text scores,
+        is_mixed flag, and evidence IDs.
+    """
+    if not per_text_scores:
+        return CompanySentimentResult(
+            company_id=company_id,
+            label="neutral",
+            positive_prob=0.0,
+            negative_prob=0.0,
+            neutral_prob=1.0,
+            evidence_ids=[],
+            is_mixed=False,
+            per_text_scores=[],
+            model_version=model_version,
+            calibration_version=calibration_version,
+        )
+
+    # Detect mixed sentiment from disagreement
+    is_mixed = _detect_mixed_from_disagreement(per_text_scores)
+
+    # Compute average probabilities across texts
+    n = len(per_text_scores)
+    avg_pos = sum(s.positive_prob for s in per_text_scores) / n
+    avg_neg = sum(s.negative_prob for s in per_text_scores) / n
+    avg_neu = sum(s.neutral_prob for s in per_text_scores) / n
+
+    # Normalize to ensure probabilities sum to 1.0
+    total = avg_pos + avg_neg + avg_neu
+    if total > 0:
+        avg_pos /= total
+        avg_neg /= total
+        avg_neu /= total
+    else:
+        avg_pos = 0.0
+        avg_neg = 0.0
+        avg_neu = 1.0
+
+    # Determine label
+    if is_mixed:
+        label = "mixed"
+    else:
+        label = _argmax_label(avg_pos, avg_neg, avg_neu)
+
+    evidence_ids = [s.evidence_id for s in per_text_scores]
+
+    return CompanySentimentResult(
+        company_id=company_id,
+        label=label,
+        positive_prob=round(avg_pos, 6),
+        negative_prob=round(avg_neg, 6),
+        neutral_prob=round(avg_neu, 6),
+        evidence_ids=evidence_ids,
+        is_mixed=is_mixed,
+        per_text_scores=per_text_scores,
+        model_version=model_version,
+        calibration_version=calibration_version,
+    )
+
+
+def _detect_mixed_from_disagreement(per_text_scores: list[TextSentiment]) -> bool:
+    """Detect mixed sentiment from evidence-group disagreement.
+
+    Returns True when at least one text has positive_prob >= threshold
+    AND at least one (different) text has negative_prob >= threshold.
+    This indicates conflicting evidence directions.
+
+    A single text cannot trigger mixed on its own (we need disagreement
+    between at least 2 texts).
+    """
+    if len(per_text_scores) < 2:
+        return False
+
+    max_pos = max(s.positive_prob for s in per_text_scores)
+    max_neg = max(s.negative_prob for s in per_text_scores)
+
+    return max_pos >= MIXED_DISAGREEMENT_THRESHOLD and max_neg >= MIXED_DISAGREEMENT_THRESHOLD
+
+
+def _argmax_label(pos: float, neg: float, neu: float) -> str:
+    """Return the label with the highest probability."""
+    if pos >= neg and pos >= neu:
+        return "positive"
+    elif neg >= pos and neg >= neu:
+        return "negative"
+    return "neutral"