fix: HuggingFace model pricing, vendor detection, cost precision (v1.9.3)

[devonakelley]

Summary

Three fixes for $0 cost display and missing model pricing in telemetry.

Fix 1: pricing.py — Missing HuggingFace models

Added pricing entries for models actively used in the routing pipeline that were returning fallback/wrong costs:

• meta-llama/llama-3.3-70b-instruct ($0.90/$0.90 per 1M)
• mistralai/mixtral-8x22b-v0.1 ($1.20/$1.20 per 1M)
• qwen/qwen2.5-72b-instruct ($0.90/$0.90 per 1M)
• deepseek/deepseek-r1 ($0.55/$2.19 per 1M)

Also added a HuggingFace fuzzy matching block to normalize_model_name() — this block was missing entirely, meaning all HF models fell through to raw string lookup and often missed.

Fix 2: openai_instr.py — Vendor detection for HF-routed models

Models like meta-llama/Llama-3.3-70B-Instruct are called via the OpenAI-compatible API but were being attributed as vendor openai, hitting the GPT-4 pricing fallback ($30/$60 per 1M) instead of correct HF pricing.

Updated _detect_vendor() to detect Llama, Mixtral, and Qwen model names and route them to the huggingface vendor for correct cost attribution.

Fix 3: collector.py — Cost precision

Cost was stored at 6 decimal places but the dashboard was rendering small values as $0.00. Rounded cost_usd to 4 decimal places in the event payload so sub-cent costs (e.g. $0.0028) display correctly instead of showing $0.

Testing

All 72 existing tests pass. No breaking changes — all fixes are additive.

Version

Bumped to 1.9.3.