fix: request streaming usage in Anthropic translation

[EleanorWho]

Summary

• Set stream_options.include_usage = true on the translated Chat Completions request when an Anthropic streaming request is forwarded, so the backend returns token usage in its final chunk.
• Extract prompt_tokens, completion_tokens, and prompt_tokens_details.cached_tokens from the upstream usage chunk and map them into the Anthropic message_delta usage fields.
• Apply the same cache subtraction logic as the non-streaming path (input_tokens = prompt_tokens - cached_tokens).

Closes #803

Test plan

• streaming_request_includes_usage_option — verifies stream_options.include_usage is injected for stream: true
• non_streaming_request_omits_stream_options — verifies no stream_options without streaming
• stream_false_omits_stream_options — verifies no stream_options for stream: false
• message_delta_usage_matches_anthropic_schema — verifies output_tokens and input_tokens are mapped
• message_delta_usage_with_cached_tokens — verifies input_tokens = prompt_tokens - cached_tokens and cache_read_input_tokens is set
• message_delta_usage_without_usage_chunk — verifies graceful default when no usage chunk is present
• all_inference_fixtures_replay — updated basic-stream fixtures for openai and vllm providers
• make test passing (all unit + integration tests)
• make lint passing (clippy + nightly fmt + machete)

🤖 Generated with Claude Code

[aslakknutsen]

"basic-stream recordings for openai and vllm now include stream_options.include_usage on the upstream request and the replay helper mirrors that. However, upstream SSE recordings still end at the finish_reason chunk with usage:null and no trailing usage-only chunk. Expected client message_delta in the openai fixture still shows output_tokens:0 and input_tokens:null, so fixture replay does not demonstrate restored billing counters end-to-end."

[EleanorWho]

Thanks for pointing out the issue. I'll re-record the OpenAI fixture.
For vLLM, is there a shared environment I can use, or should the vLLM fixture stay as-is for now? It still passes and covers the fallback when the backend doesn't return usage.

[praxis-bot]

PR Review

Summary: Injects stream_options.include_usage on translated streaming requests and maps the upstream usage chunk into Anthropic message_delta usage fields.

Overall: Well-implemented with thorough test coverage. The streaming usage mapping is correct, including the cache subtraction logic using saturating_sub. One doc comment was inadvertently split.

Severity	Count
Critical	0
Large	0
Medium	1

Findings without inline placement

(none)

[praxis-bot]

[Medium] The convert_stream function was inserted between lines 1 and 3 of map_parameters' existing doc comment, splitting it. As a result, convert_stream now has map_parameters' summary ("Map Anthropic parameters to Chat Completions-compatible equivalents.") as its first doc line, which is inaccurate, and map_parameters lost its summary (now starting with the top_k detail paragraph).

Replace convert_stream's doc with an accurate single-line summary and restore map_parameters' summary:

/// Copy `stream` and inject `stream_options.include_usage` when enabled.
fn convert_stream(...)

/// Map Anthropic parameters to Chat Completions-compatible equivalents.
///
/// `top_k` has no standard Chat Completions equivalent but is preserved
/// as an extra body parameter for backends that support it
/// (e.g. vLLM).
fn map_parameters(...)

[leseb]

@EleanorWho is this addressed?

[EleanorWho]

It's addressed in the latest commit.

[praxis-bot-app]

Unsigned commits: 3fd7ce6, 9663554. Please sign your commits.

[praxis-bot-app]

AI tool authorship detected:

• 3fd7ce6: Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
• 9663554: Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Sorry, this project does not accept commits authored by tools as valid.
Commits need to be authored by and signed-off by the human(s) responsible for the PR, with their name and contact.

[leseb]

@EleanorWho please rebase

[EleanorWho]

@leseb Hi Seb, can you please take a look at the recording fixture in the latest commit? The previous OpenAI recording had "kind": "imported" with a source_id pointing to a Python test, but my re-recording using cargo xtask record-inference produced "kind": "live" with source_id: null.
I'm not sure if this provenance change is acceptable or if I should re-record through the import path to preserve the original format.

[leseb]

Two actionable fixture-governance issues remain:

The vLLM live fixture’s request was manually updated without re-recording its response. It still lacks the trailing usage chunk, although vLLM emits one for include_usage: true. Re-record it truthfully. vLLM documentation

coverage.yaml needs explicit feature inventory for the new streaming usage request/response transformation, per AGENTS.md, followed by README regeneration.

[leseb]

@leseb Hi Seb, can you please take a look at the recording fixture in the latest commit? The previous OpenAI recording had "kind": "imported" with a source_id pointing to a Python test, but my re-recording using cargo xtask record-inference produced "kind": "live" with source_id: null. I'm not sure if this provenance change is acceptable or if I should re-record through the import path to preserve the original format.

The provenance change is correct. Since this fixture was captured using cargo.