fix(catalog): semantic search works again, and says why when it cannot

[jrosskopf]

The audit's last group: the catalog commands. Two findings, one of them a live outage.

catalog search --mode vss|hybrid was broken by something outside this repo

Both failed with a bare HTTP 404 from the embedding provider — with a valid API key. The cause: Google retired text-embedding-004, so the endpoint we call no longer exists. catalog build --embed was broken the same way.

$ erpl-adt catalog search catalog.duckdb "procurement value" --mode hybrid
Error: GeminiEmbeddingProvider::EmbedText (HTTP 404)      # before
  [ABAP] DTEL/DE AD_PROCUREMENT_DOC  (score 0.0163934)    # after

gemini-embedding-001 replaces it, with two details that matter:

• It returns 3072 dimensions by default where the catalog schema stores FLOAT[768], so the request asks for 768 explicitly — every existing catalog stays readable.
• Its truncated vectors are not unit length, unlike the ones text-embedding-004 returned, so they are normalised before storage. Without that, cosine similarity would be inconsistent between rows written before and after this change.

The model name now comes from ERPL_ADT_GEMINI_EMBED_MODEL when set, so the next retirement is a configuration change rather than a release, and a 404 from the provider names the unavailable model instead of surfacing a bare status.

A quieter one in the same command

A catalog built without --embed holds no vectors, and a semantic search over none returned an empty list — indistinguishable from "nothing matched", the same trap as the ADT commands in #52. It now says the catalog holds no embeddings and points at --mode fts or a rebuild.

Verified

Against a 100 MB catalog on this machine: hybrid search returns ranked hits, and --mode vss against that (unembedded) catalog explains itself. 1349 unit tests pass.

A correction worth recording

While fixing this I first "explained" the build failure as a GCC -O3 quirk and silenced it with inline. That was wrong: the compiler was right — a member ModelName() in the header shadowed the free function I had added, so my new model name was never used and the old one was still being sent. The member is now defined in the .cpp with the configurable logic, and the bogus comment is gone.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}