fix(zen): stop double-counting reasoning_tokens in oa-compat usage - #24367

Summary

Fixes #24268.

The OpenAI chat-completions usage spec says completion_tokens already includes completion_tokens_details.reasoning_tokens. Zen's oaCompatHelper.normalizeUsage was reporting outputTokens = completion_tokens and reasoningTokens = reasoning_tokens without subtracting, so downstream calculateCost (which bills outputCost + reasoningCost separately at the same cost.output rate) double-billed reasoning.

The reporter's evidence:

• API: prompt_tokens=22, completion_tokens=77, completion_tokens_details.reasoning_tokens=78 for a single "Hi"
• Local opencode db: output: 436, reasoning: 790 for a real session (correctly excludes reasoning from output)
• Zen Usage console (broken): output: 1226, reasoning: 790, total: 2016 — i.e. completion_tokens (1226) + reasoning_tokens (790), double-counting
• Cost was being computed off the inflated 2016 total

Fix

Mirror openaiHelper.normalizeUsage from the OpenAI Responses helper (openai.ts) and subtract reasoning_tokens from completion_tokens before returning. Clamp at 0 because some OA-compatible providers report reasoning_tokens > completion_tokens (e.g. Moonshot Kimi K2.6 returning reasoning=78, completion=77).

This keeps outputTokens + reasoningTokens === completion_tokens (or close to it under provider quirks), which is the invariant calculateCost depends on to get the same total bill the upstream API uses.

Tests

Adds packages/console/app/test/zen-usage.test.ts with 4 cases that lock in the wire-level invariant for both helpers:

• Reporter's real session: completion=1226, reasoning=790 → outputTokens=436, reasoningTokens=790, sum=1226 ✅
• Reporter's "Hi" with inverted ratio: completion=77, reasoning=78 → outputTokens=0, reasoningTokens=78 (clamped) ✅
• No reasoning at all: outputTokens left untouched ✅
• Parity with openaiHelper.normalizeUsage for the same logical usage ✅

Test plan

• bun test in packages/console/app/ — 7/7 pass (was 3/7 with the new tests written against the old code, all 4 new tests fail before the fix and pass after)
• bun typecheck from repo root — all 13 tasks successful
• No new lint errors on the modified file

Notes

• I deliberately did not also touch openaiHelper.normalizeUsage (which already subtracts but does not clamp), since the OpenAI Responses API doesn't exhibit the reasoning > completion quirk and changing it is out of scope.
• The bookkeeping nuance the reporter raised in the "Extended Evidence" section (cases where reasoning > completion cause weird per-message rows) becomes a clean outputTokens=0, reasoningTokens=N row instead of a negative outputTokens after this fix. Total cost ends up matching the upstream API's completion_tokens * rate to within one token of rounding.