Extend datacortex from retrieval-only to predictive embeddings

[plur9]

Context

Datacortex stores semantic embeddings for similarity search. Kumo's relational pretraining produces predictive embeddings — vectors learned from outcomes, not from text similarity. Same storage substrate, different learning objective.

Reference: https://kumo.ai/company/news/low-latency-high-throughput-predictions-with-kumorfm-2-fine-tuning/

Problem

Today datacortex answers "what's similar to this?" but not "what's about to happen to this?" — even though the data to answer the second question (outcomes: task completion, contact response, venture hypothesis confirmation) is sitting in Datacore.

Proposal

Add a second embedding kind to the datacortex store: outcome-supervised vectors trained on labeled events.

Examples:

• Task embedding trained to predict "was this completed within deadline?"
• Contact embedding trained to predict "will this contact respond in 30 days?"
• Venture-hypothesis embedding trained to predict "did this hypothesis hold?"

Reuse the existing datacortex vector backend; add a model field per embedding. Retrieval can still query by similarity, but ranked by predicted outcome.

Acceptance

• Datacortex schema accepts multiple embedding kinds per entity
• At least one outcome-supervised embedding shipped (recommend: task completion likelihood)
• Datacortex queries can rank by predicted outcome, not just similarity
• Documented in datacortex module README

Depends on

• "Relational router model" issue (model providing the outcome predictions)