fix(memory): rebuild vec tables on embedder dimension change

[glitch-ux]

Closes #137.

Problem

decisions_vec / turn_summaries_vec are created at a hardcoded float[384] (_VEC_DIM) and never rebuilt when the project's embedder has a different dimension. _write_vec_row swallows the resulting dimension-mismatch OperationalError at debug, so every semantic-memory write is silently dropped and session_recall returns nothing, permanently — no error surfaced. This hits the common zero-dependency path: without the [local] extra, cce falls back to Ollama, whose default nomic-embed-text is 768-dim.

The comment at memory/db.py:30 already pointed at a _ensure_vec_dim rebuild — but it was never implemented (grep matched only the comment). The code index solves the same problem via VectorStore._ensure_vec_table; this brings the memory subsystem to parity.

Fix

• Add _ensure_vec_dim(conn, dim) (+ _declared_vec_dim): when the active embedder's dimension differs from the vec tables', drop and recreate both *_vec tables at the new dimension. The decisions / turn_summaries source rows are untouched — backfill_vec_tables re-embeds them.
• Call it from the record_*_vec writers (self-heals on the next write) and reconcile up front in backfill_vec_tables (runs at cce serve startup, so an already-populated index is repaired too — the NOT EXISTS filter alone would miss that case).

Both vec tables always share the embedder's dimension, so they're rebuilt together. No change to the source schema, migrations, or the 384-dim default path.

Testing

• New tests/memory/test_vec_dim.py: rebuild-on-larger-dim + row persists + searchable, matching-dim is a no-op (no data loss), and backfill-after-swap repopulates. These fail on main (silent drop / missing function) and pass here.
• Full tests/memory/ suite green (362 passed) on Python 3.13 with the sqlite-vec extension available.
• ruff check src/ tests/ clean.
• The repro from Memory vec tables hardcode dim=384 and are never rebuilt on embedder change → semantic recall silently dies #137 now returns vec rows: 1 / search: [1].