Declare a single corpus embedding-space contract in the embed route + re-embed migration on backend switch (#602/#82 follow-up)

[mdheller]

Follow-up to #605 (fix/embedding-space-pin-602), which pinned the query path and made a dim mismatch fail loud in lib/memory retrieval. That stops the silent degrade-to-lexical, but the root asymmetry upstream remains and is out of scope for the surgical fix:

• app/api/embed/route.ts produces vectors at different dims depending on backend: OpenAI text-embedding-3-small with dimensions: 512 vs local nomic-embed-text at 768. There is no single declared CORPUS_EMBED_DIM the route pins to.
• Consequence: switching backends (add/remove an OpenAI key, change NOETICA_EMBED_MODEL) silently produces a corpus embedded in a different space than later queries. fix(memory): pin corpus embedding dim on the query path — fail loud, never degrade to lexical (#602/#82) #605 now catches this at query time by raising — good (fail-closed), but the user experience is a hard error rather than a repaired corpus.

Beyond-the-fix work to make the space a real contract on ALL paths:

1. Declare one authoritative CORPUS_EMBED_DIM (and model id) as estate config; the embed route stamps the produced space (model + dim) into the response / stored entry metadata.
2. Persist the embedding space alongside each stored vector (MemoryEntry) so a mismatch is detectable without inference from the first entry.
3. Re-embed migration: when the active backend's space differs from the stored corpus space, re-embed (or refuse writes into a mixed store) instead of accumulating mixed-dim entries.
4. Mirror the same PinnedSpace check in any OTHER consumer of /api/embed (graph/sae/tune paths) — the estate rule is pin on ALL paths incl. callbacks.

Refs #605, #602, prophet-workspace#82, prophet-workspace#76. cc @mdheller