feat(storage): add complete embedding-space identity

[ohdearquant]

Stack and governance gate

• Stacked on feat(retrieval): add bounded ranked-prefix materialization #2015 (codex/adr-160-phase-5-ranked-prefix).
• Implements ADR-160 Phase 6 only.
• ADR-160 was accepted on main in 135c3d46 (docs(adr): mark ADR-160 accepted #2005). Keep this PR draft until the lower stacked phases and the Phase 4 operational rollout gates are satisfied and Claude completes review.

What changed

• Add a shared, closed EmbeddingSpaceIdentity / EmbeddingSpaceKey contract in khive-storage, derived from a complete protocol fingerprint plus dimensions rather than a display/model name.
• Validate borrowed identity inputs and preflight the 128-byte derived-key envelope before allocation; keep the key field private and provide no unchecked constructor or deserializer.
• Replace runtime's pack-local NamedVectorIdentity seam with vectors_for_embedding_space, binding the complete key to immutable SQLite vector-table geometry and rejecting conflicting first registration.
• Migrate moodboard's descriptor path to the shared identity while preserving its existing canonical JSON, SHA-256 fingerprint, public response shape, and table key.
• Capture the identity once alongside the loaded vision model and carry it through inference to storage, avoiding per-request descriptor reconstruction and hashing.
• Add a cross-language Python descriptor golden to the repository smoke gate, plus exact Rust response/table and field-mutation goldens.
• Document the intentional source break and the table-bound/default-namespace behavior accurately.

Resource and failure-safety properties

• Protocol/model/key-prefix inputs are bounded; model revision input is capped at 4096 bytes.
• Oversized prefixes fail from borrowed input before a second large allocation or derived-key formatting.
• Dimensions are restricted to 1..=8192; derived table keys are restricted to 128 bytes.
• Existing-table model/geometry conflicts fail closed; concurrent first registration has one immutable winner.
• Production moodboard requests reuse the identity cached at model load.

Regression guards

• Storage tests pin grammar/bounds, key derivation, exact ceiling behavior, a 1 MiB borrowed-prefix refusal, and the private-key compile-fail boundary.
• Runtime tests pin model/geometry conflict refusal, planted-table mismatch, concurrent first-bind behavior, fingerprint isolation, namespace-scoped rows over a shared table, and persistent restart retrieval.
• Moodboard tests pin canonical bytes/fingerprint/key, the complete 12-field response, exact physical table selection, and a mutation matrix covering every vector-affecting descriptor field.
• The independent Python standard-library golden reconstructs the canonical descriptor and asserts byte-, SHA-, and key-level parity.

Validation

• RUSTC_WRAPPER= cargo check --workspace -j4
• RUSTC_WRAPPER= cargo clippy --workspace --all-targets -j4 -- -D warnings
• RUSTC_WRAPPER= cargo test --workspace -j4
• python3 tests/test_moodboard_descriptor_golden.py
• cargo fmt --all -- --check
• deno fmt --check on all eight modified Markdown files
• bash -n scripts/ci.sh
• git diff --check

Independent identity-contract, callsite/docs, registry-boundary, and resource reviews found no implementation Blocker/High/Medium issue.

Explicitly deferred to Phase 7

This PR does not migrate the text-provider registry, ANN/cache identities, or embedding-model schema. Accepted ADR-160 now fixes the configured lineage-slot and legacy-row mapping; Phase 7 implements that cutover and must also source a complete attested resolved-artifact identity covering provider/version, query/document transforms, pooling/execution semantics, and checkpoint digests. Those concerns remain outside this Phase 6 boundary.