`assertUsageDisjointness` covers 6 of 16 AI providers — xAI and llamacpp-server map more usage fields than Ollama, which is covered

[sroussey]

What

assertUsageDisjointness (packages/test/src/contract/ai-provider/assertions/usageNormalization.ts:91) is the assertion the 2026-08-17 review credited as making Usage.input disjoint "with a generic contract test … so a new provider inherits the contract by construction". It asserts input + cached + cacheWrite + output === total for every case of every provider block.

It is invoked through usageNormalizationBlock, which packages/test/src/test/ai-provider/ProviderUsageNormalization.test.ts calls seven times, for six providers:

$ grep -n "provider:" packages/test/src/test/ai-provider/ProviderUsageNormalization.test.ts
39:  provider: "OpenAI (chat-completions shape)",
86:  provider: "OpenAI (Responses shape)",
115: provider: "Anthropic",
163: provider: "Gemini",
192: provider: "Ollama",
220: provider: "DeepSeek",
249: provider: "OpenRouter",

Not covered, with how much usage mapping each actually carries in src:

$ for p in providers/*/; do echo "$(basename $p) $(grep -ric usage $p/src | awk -F: '{s+=$2} END{print s+0}')"; done
xai 61                      ← not covered
llamacpp-server 58          ← not covered
huggingface-inference 44    ← not covered
node-llama-cpp 36           ← not covered
huggingface-transformers 19 ← not covered
chrome-ai 2                 ← not covered
cactus 0 · tf-mediapipe 0 · stable-diffusion-server 0 · mlx 0
--- covered: openrouter 79 · anthropic 73 · deepseek 72 · google-gemini 66 · ollama 61 · openai 54

xai (61) and llamacpp-server (58) each touch more usage code than ollama (61, tied) — and ollama got a block.

Why it matters

• xAI is OpenAI-compatible in shape, which is exactly the assumption the disjointness contract exists to stop people making silently: OpenAI's prompt_tokens includes cached and Anthropic's input_tokens excludes it, and the whole point of the axis is that the same-looking field means different things per vendor. Xai_Usage.ts is unasserted on that question.
• llamacpp-server, node-llama-cpp and huggingface-transformers report locally-computed usage with no vendor to check against, which is the case most likely to double-count. createDecodeUsageReporter and createEstimatedOutputUsageReporter both feed cumulative snapshots; nothing asserts the snapshot is disjoint.
• Four providers report no usage at all (cactus, tf-mediapipe, stable-diffusion-server, mlx), which is its own finding: stable-diffusion-server has had grep -rc usage src → 0 for six consecutive review cycles behind a 125-line README that does not mention it.

Proposed fix

1. Add blocks for xai and llamacpp-server first — highest mapping surface, lowest effort, and the fixtures are the same shape as the OpenAI and Ollama blocks already in the file.
2. Then node-llama-cpp and huggingface-transformers, whose numbers are computed rather than received.
3. Make the omission visible rather than silent: drive ProviderUsageNormalization.test.ts off a record keyed by provider name with an explicit usageMapping: "none" entry for the four that report nothing, so a provider that starts reporting usage without a block is a type error. Same shape as PROVIDER_CASES / PRICED_CASES in inferAdvertisesRegistered.test.ts, which already pins the analogous gap for capabilities and pricing.
4. For the four with zero usage, either wire it or say so in the README — stable-diffusion-server in particular.

Found during the 2026-09-14 review of providers/. Verified against origin/main @ 2d36880.