fix: reject malformed embedding indices before lossy coercion

[DivyamTalwar]

@haolpku

Summary

The OpenAI-compatible embedding provider now validates response indices before sorting or coercing them.

Root cause

The old path called int(row["index"]), allowing fractional values such as 0.9 and 1.9 to become 0 and 1. The completeness check then accepted a malformed response. Python booleans also passed the numeric coordinate check because bool subclasses int.

Change

• Require response indices to be actual JSON integers and reject booleans, fractional values, and numeric strings.
• Validate range, uniqueness, and completeness only after strict type validation.
• Reject boolean vector coordinates.
• Normalize non-object response bodies and malformed rows through ExternalServiceError.
• Keep valid out-of-order rows, dimension detection, HTTP retry policy, and model/batching behavior unchanged.

Regression coverage

The focused tests cover fractional, boolean, string, negative, duplicate, and out-of-range indices; boolean coordinates; valid reordering; and the no-retry rule for deterministic schema errors.

Verification: python -m pytest -q — 202 passing, 5 optional SDK tests skipped.

No network service or embedding quality claim is made; this is response-shape validation only.

Please review and merge if CI is green.

[DivyamTalwar]

@haolpku — Could you please review the strict embedding-response validation fix and merge it if CI confirms the full suite and compatibility expectations?