fix(azure): return embedding token usage instead of only logging it

[bobmcwhirter]

Azure's embedding client reports zero token usage to callers, because it never overrides EmbeddingModel::embed_texts_with_usage.

The data isn't missing — it's discarded. embed_texts already parses response.usage, hands it to tracing::info!, and then drops it on the floor:

// crates/rig-core/src/providers/azure.rs
ApiResponse::Ok(response) => {
    tracing::info!(target: "rig", "Azure embedding token usage: {}", response.usage);
    ...
    Ok(response.data.into_iter()...collect())   // usage gone
}

Since the trait's default embed_texts_with_usage returns Usage::default() (embeddings/embedding.rs), anything reading EmbeddingResponse::usage gets zeroes for every Azure embedding request.

The change

Restructured exactly the way providers/openai/embedding.rs already does it: embed_texts_with_usage does the work and returns the usage, embed_texts delegates and drops it. 14 insertions, 2 deletions, one file.

No new mapping code was needed — impl GetTokenUsage for Usage already exists in azure.rs and was simply unused for embeddings. The tracing::info! line is retained, so anything watching logs is unaffected.

How it was found, and verified

Measured against a real Azure OpenAI deployment: 100 embedding requests, 0 tokens recorded. After this change the same workload reports the provider's actual prompt_tokens. In our case it meant every embedding cost figure downstream read $0.00 on real spend, which is how we noticed.

Test

An #[ignore]d live test (tests/providers/azure/embeddings.rs), which is this provider's existing convention — there is no tests/cassettes/azure/ fixture set or with_azure_cassette helper to extend, and both current azure tests are #[ignore]-gated on AZURE_OPENAI_API_KEY. CONTRIBUTING lists ignored live tests as acceptable where cassette replay is unsuitable.

I deliberately did not hand-author or record a cassette: recording one against our deployment would have committed our Azure resource and deployment identifiers into this repository. Happy to add a cassette if you'd prefer one recorded from your own fixtures — say the word and I'll follow up.

cargo fmt --check clean; cargo check -p rig-core and cargo test --test azure --no-run both build.

Related, but deliberately out of scope

Only 2 of the 8 provider files implementing embed_texts override embed_texts_with_usage (openai, voyageai). The other six inherit the zero-returning default, so they very likely share this defect — I've kept this PR to azure to stay small and reviewable per CONTRIBUTING, rather than touching providers I can't test.

Worth considering separately: a default implementation that silently returns plausible-but-wrong zeros is easy to inherit by accident, and it took a production cost discrepancy for us to notice. Making embed_texts_with_usage a required method, or having the default return something callers can distinguish from a real measurement, would surface this class of bug at compile time instead. Glad to open an issue for that discussion if it's useful.

[gold-silver-copper]

Provider pipelines have been greatly consolidated, this should be fixed on main now. Please open a new issue if the bug persists. Thank you 🙏