A retrieval mode the code cannot perform, advertised in three places

A retrieval mode the code cannot perform, advertised in three places Caught while drafting a post about the search story, by checking the claim before writing it down: **nothing in this workspace generates an embedding.** No SigLIP, no CLIP, no hosted embedding call, and no statement anywhere writes a vector column. The `pgvector` columns exist and the planner can express a semantic clause; nothing fills them. The claim was live in three places, on a repository public since this morning: - README — "Lexical, faceted, and **semantic** retrieval share an explainable query model" - `dam-search`'s module header — "Tantivy index, **pgvector, hybrid query planner**" - and the crate's manifest description, because those descriptions were generated *from* the module headers a few commits ago. An inaccurate header propagated into package metadata without anybody restating it, which is the part worth remembering. All three now say what is true: lexical and faceted retrieval share one query model, and the vector half is schema and planner awaiting the model-distribution decision TASKS.md already carries as open. That decision is about shipping a multi-gigabyte ONNX file rather than writing a query — which is why it stayed open, and why it was so easy to describe as though it were finished. The module header now carries that reason, because it is the string the manifest copies. `.gitignore` learns about `.terraform/` here as well. It was added on the unmerged branch that introduced the Terraform, so `git add -A` on this branch swept in a 780MB provider binary and GitHub rejected the push — an ignore rule is only doing its job on the branch it exists on.