fix(stream): correct cached input accounting

[IfkumRfnl]

Summary

• Use inputTokenDetails.noCacheTokens for uncached input usage.
• Fall back to subtracting cache reads and writes from total input.
• Add regression coverage using the usage shape returned by Command Code.

Command Code's inputTokens includes cached tokens. Treating it as uncached input caused cache reads to be counted twice in pi's usage totals, cost estimates, and context accounting.

Testing

• npm run typecheck
• npm run test:unit
• npm run test:stream
• npm run format:check
• Manually verified with deepseek/deepseek-v4-flash

Fixes #30

[patlux]

Verified the token semantics against the Vercel AI SDK usage schema and the provider implementations used by Command Code. The fix correctly separates uncached input, cache reads, and cache writes. Added regression coverage for the fallback without noCacheTokens and for cacheWriteTokens; local tests and CI are green.

[patlux]

Thank you @IfkumRfnl for reporting this with a concrete API example and providing the fix! I independently verified the usage semantics against the Vercel AI SDK and the provider implementations used by Command Code. I also added coverage for the fallback path and cache-write accounting. Everything is green.