Fix OpenAI embedding requests and validate embedding model coverage

Current review — 2026-10-20

The prior investigation PR phaseoteam/Phaseo#1609 is closed unmerged. Current main has OpenAI/Google embeddings executors and shape/protocol/coverage tests, but no request-specific evidence here establishes the reported production failure was resolved. Remaining: reproduce using a current supported OpenAI model, trace eligibility/pricing/translation, and record single/batch/dimensions/encoding/usage plus a non-OpenAI regression check.

---

Problem

OpenAI embedding requests currently appear not to work. The failure may involve model identity, provider routing, request translation, response normalization, or metadata.

Investigation scope

• Reproduce with supported OpenAI embedding models.
• Trace catalog/API-model identity and provider eligibility.
• Validate request translation for single and batched inputs.
• Validate dimensions, encoding format, usage, timing, and response normalization.
• Test additional embedding models/providers so the fix is not OpenAI-specific where shared infrastructure is involved.

Acceptance criteria

• Supported OpenAI embedding requests succeed through the intended route.
• Single-string, array, and representative dimensions/encoding inputs behave according to the API contract.
• Usage and response metadata are populated correctly.
• Invalid or unsupported inputs return actionable errors.
• At least one non-OpenAI embedding route is checked for regression where coverage permits.
• Regression tests protect the identified root cause.