fix(rag): make OCI embedding dimension + input_type deterministic (closes #292)

[fede-kamel]

Closes #292.

The report

RAG queries against a cohere.embed-v4.0-indexed Oracle table failed with:

ORA-51803: Vector dimension count must match the dimension count specified in
the column definition (expected 1536 dimensions, specified 1024 dimensions).

The original hypothesis was that OCIEmbeddings.embed_query() hardcoding
input_type="SEARCH_QUERY" caused v4 to return 1024-dim query vectors against a
1536-dim (SEARCH_DOCUMENT-indexed) table.

What the investigation actually found

I reproduced against live OCI Generative AI (us-chicago-1) on two separate tenancies.
input_type does not change the output dimension — that hypothesis is wrong:

model	SEARCH_DOCUMENT	SEARCH_QUERY
cohere.embed-v4.0	1536	1536
cohere.embed-english-v3.0	1024	1024
cohere.embed-multilingual-v3.0	1024	1024
cohere.embed-english-light-v3.0	384	384

The real lever is Cohere v4's Matryoshka output_dimensions, which Locus never
set. v4 returns exactly the requested size, and OCI's server-side default applies when
it's omitted:

v4 default (no output_dimensions): 1536
v4 output_dimensions=256:  256
v4 output_dimensions=512:  512
v4 output_dimensions=1024: 1024   ← the reported 1024
v4 output_dimensions=1536: 1536

So the ORA-51803 happens when one path (an ingestion pipeline, an older server-side
default, or a foreign OracleVS writer) lands at one output_dimensions and another
path lands at a different one. Because Locus left the dimension to an implicit OCI
default, the stored column and query vectors could silently diverge. input_type was a
red herring — but the code did genuinely hardcode it, so that's fixed too.

Fix

Primary — deterministic dimension (output_dimensions):

• Add OCIEmbeddingConfig.output_dimensions (default None). It's forwarded to
  EmbedTextDetails on every embed path through a shared _build_embed_details helper,
  and only when set, so non-Matryoshka models (v3, fixed dims) never receive the
  field.
• config.dimension now prefers output_dimensions, so the vector store sizes its
  column to exactly the embedding dimension being produced.

Secondary — honor input_type (the original hardcode + cleanup):

• embed_query() no longer hardcodes SEARCH_QUERY; it reads a new
  OCIEmbeddingConfig.query_input_type (default SEARCH_QUERY).
• embed_documents() reads input_type from config instead of hardcoding
  SEARCH_DOCUMENT.
• Removed the dead original_type local and the incorrect "Can't modify frozen config"
  comment in embed_query (the config is a plain BaseModel, never frozen).

Tests

Unit — tests/unit/test_rag_embeddings_oci.py:

• output_dimensions omitted by default (never attached for non-Matryoshka models)
• output_dimensions forwarded on all three paths (embed/embed_query/embed_documents)
• config.dimension reflects output_dimensions (wins over the model hint)
• query_input_type / input_type are honored, not hardcoded
• config defaults assert the two new fields

Full unit suite: 5023 passed, 0 failed.

Integration — tests/integration/rag/test_oci_embeddings.py: 9 passed against live
OCI GenAI.

Live verification of the fix (real cohere.embed-v4.0, us-chicago-1):

config	embed_documents	embed_query	config.dimension
default (unpinned)	1536	1536	1536
output_dimensions=1024	1024	1024	1024

Both paths and the column dimension stay in lockstep — the ORA-51803 divergence can no
longer arise from an implicit server default.