fix(stream): estimate input tokens when upstream reports prompt_tokens=0

[adivekar-utexas]

Summary

Some providers (notably Ollama Cloud) return prompt_tokens: 0 in their streaming usage-only chunk because they do not count input tokens at all. The streaming handler accepted this as valid usage (since completion_tokens > 0), so the estimation fallback never ran, and x-omniroute-tokens-in=0 was emitted — showing "Total In: 0" on the dashboard despite real input.

Root Cause

In open-sse/utils/stream.ts, the passthrough empty-choices handler:

1. extractUsage() returns {prompt_tokens: 0, completion_tokens: N} — correctly parsing Ollama's upstream
2. hasValidUsage() returns true because completion_tokens > 0
3. The estimation fallback at flush time (if (!hasValidUsage(usage))) never runs
4. prompt_tokens: 0 propagates to x-omniroute-tokens-in and the dashboard

Fix

When the usage-only chunk has prompt_tokens: 0 but non-zero completion_tokens, estimate the actual input token count from the request body and patch the usage object. Uses the existing estimateUsage(body, totalContentLength, ...) function.

Test Plan

• New test: createSSEStream passthrough estimates input tokens when upstream reports prompt_tokens=0 (Ollama Cloud pattern) — sends an Ollama-style stream with prompt_tokens: 0 / completion_tokens: 10 and asserts the output has prompt_tokens > 0 and total_tokens == prompt_tokens + completion_tokens
• ✅ Test fails without fix: prompt_tokens should be estimated (> 0), got 0
• ✅ Test passes with fix: in=2041 estimated from request body
• All 52 stream-utils tests pass (51 existing + 1 new)
• Pre-commit hooks: prettier, eslint, any-budget, tracked-artifacts all pass cleanly

Made with Cursor

[diegosouzapw]

Obrigado, @adivekar-utexas! 🙌 Reconstruí sobre o tip da release/v3.8.34 (só o commit da feature; o estimador de input tokens é o mesmo já usado em 4 outros pontos do stream.ts, sem risco de double-count), teste verde. Mergeando.