fix(llm): decode Modal reasoning token usage

[brynary]

Summary

• decode Modal’s top-level usage.reasoning_tokens field
• prefer the standard nested completion_tokens_details.reasoning_tokens field when both spellings are present
• cover both streaming and non-streaming Chat Completions usage payloads

Modal reports completion_tokens as an inclusive total and sends the reasoning subset beside it. Fabro previously discarded that sibling field, leaving all completion tokens in the visible-output bucket and reporting zero reasoning tokens.

Verification

• cargo nextest run -p fabro-llm — 671 passed, 24 skipped
• cargo +nightly-2026-04-14 fmt --check --all
• cargo +nightly-2026-04-14 clippy -p fabro-llm --all-targets -- -D warnings
• git diff --check

[Copilot]

Pull request overview

This PR updates the OpenAI-compatible wire codec in fabro-llm to correctly account for Modal’s reasoning-token reporting shape, so downstream token accounting matches the provider’s semantics (reasoning tokens separated from visible output tokens).

Changes:

• Extend ApiUsage to decode Modal’s top-level usage.reasoning_tokens.
• Normalize reasoning tokens by preferring completion_tokens_details.reasoning_tokens when present, falling back to the top-level field.
• Add focused unit tests covering both non-streaming and streaming usage payloads for Modal.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[brynary]

Merged main (resolved a conflict in wire.rs with the DeepSeek prompt_cache_hit_tokens work, which landed in the same function) and ran a simplification pass over the diff.

Production change is unchanged — a review of reuse, quality, and efficiency found no issues with it. Notably: no helper extraction for the repeated .as_ref().and_then(...).or(...) shape, since each of the three reads a different field, so a helper would need a closure per call site and end up longer. Field name and .or() also left as-is, matching the DeepSeek fallback next to it.

Cleanups:

• Collapsed the two envelope tests into focused ApiUsage tests. They were one test paid for twice — ApiResponse::usage and StreamChunk::usage are both Option<ApiUsage>, so the envelope cannot change the result. Envelope-level usage decoding is already covered by stream_chunk_usage_parsing. The payloads also carried fields ApiUsage never deserializes (total_tokens, id, model, choices) plus a cached_tokens split unrelated to the fix.
• Renamed to the token_counts_* convention used by the neighboring tests and regrouped them so the usage tests stay contiguous.
• Added a test pinning the precedence rule this PR introduces: nested detail wins over the flat spelling, and an empty completion_tokens_details still falls back to it. That behavior was previously untested. Verified both new tests fail when .or(self.reasoning_tokens) is removed.
• Documented the precedence on token_counts, which covered disjoint bucketing but not the alternate spellings.
• Reverted the unrelated cost doc edit that dropped the concrete OpenRouter reference.

Verification: cargo nextest run -p fabro-llm (690 passed, 26 skipped), fmt --check --all, clippy -p fabro-llm --all-targets -- -D warnings, git diff --check.

One out-of-scope finding for a possible follow-up: codec/openai_compatible/response.rs:81 parses the whole response body a second time into a serde_json::Value for Response.raw, which appears to have no production reader (only two fabro-llm tests). Not touched here.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.