refactor: Centralize embedding model constant for standalone scripts

[NOVA-Openclaw]

Problem

Every standalone embedding script hardcodes its own embedding model name, and they drift from the ecosystem standard (snowflake-arctic-embed2, 1024 dims). Mixed-model vectors silently corrupt the semantic index.

Evidence this is a recurring failure class, not a one-off:

• Lessons 285, 321, 537 in the lessons table all describe embedding-model drift incidents
• 2026-07-26: embed-blog-posts.py was found still using mxbai — fixed by hand, again
• chore: Unify all embedding systems on snowflake-arctic-embed2 #223 unified the ecosystem on snowflake-arctic-embed2, but enforced it by editing each script individually — nothing prevents regression

Proposed fix

1. Single shared source of truth for the embedding model name + dimensions — e.g. a config file (~/.config/nova-mind/embedding.conf or a table in nova_memory) or an env var, read by all nova-mind embedding scripts at startup.
2. Grep audit of existing scripts for hardcoded EMBEDDING_MODEL / model-name strings; migrate each to the shared constant.
3. Optional hardening: scripts fail loudly (refuse to write) if the model they resolve doesn't match the dimension recorded in memory_embeddings, instead of silently inserting mismatched vectors.

Context

• Promotion candidate from the lessons→enforcement pipeline: 3+ related lessons means the written rule failed and the system should make the mistake unrepresentable.
• Tracked as task feat(schema): provenance grading — assertion intent, S×D×V axes, entity credibility (#504) #505 in nova_memory; related task bug: workflow_runs closed with status='completed' but NULL completed_at — no schema enforcement #503 (blog_post embedding provenance audit).
• Related: chore: Unify all embedding systems on snowflake-arctic-embed2 #223 (closed — model unification), Add remaining tables to memory-maintenance.py embedding coverage #289 (embedding coverage).