fix: route embeddings through JSON parser so cost and tokens are captured

[kyto-agent]

Problem

Embedding model requests were being forced through the SSE streaming parser even though the /v1/embeddings endpoint returns plain JSON (not Server-Sent Events). This caused resolveUsage() to never receive actual data, so every embedding request was logged with:

• cost: "0" → dashboard shows "Free"
• promptTokens: 0, totalTokens: 0 → dashboard shows "0 in 0 out"
• Embedding models filtered out of the global Usage by Model table (due to the totalTokens > 0 HAVING clause)

Additionally, estimateUpstreamCost() only called fetchLanguageModels(), so embedding models were never found and every embedding call fell back to a flat $0.05 reservation regardless of actual input size.

Fix

1. Route embeddings through the JSON branch (general.ts)

Changed the condition from:

if (!body.stream && endpoint !== "embeddings")

to:

if (!body.stream || endpoint === "embeddings")

Embeddings always return plain JSON, so they need the non-streaming branch where resolveUsage(data) can extract usage.cost and usage.total_tokens from the response.

2. Use the right model fetcher for cost estimation (shared.ts)

estimateUpstreamCost() now accepts an endpoint parameter and calls fetchEmbeddingModels() for embedding requests. For embeddings, only input tokens are estimated (no completion side). This gives an accurate reservation instead of the flat $0.05 fallback.

3. Include zero-token models in stats when they have cost (stats.ts)

Added OR COALESCE(SUM(cost), 0) > 0 to the HAVING clause so models with non-zero cost but potentially zero tokens (some embedding/image edge cases) still appear in "Usage by Model".

Files changed

• src/routes/proxy/v1/general.ts — routing condition + pass endpoint to cost estimator
• src/routes/proxy/shared.ts — import fetchEmbeddingModels, branch cost estimation by endpoint
• src/lib/stats.ts — relax HAVING clause to include costly zero-token models