Token count fields don't distinguish estimated vs actual values

[piotrekno1]

Problem

TextInputMeta.tokenCount and GenerationOutputMeta.tokensOut are emitted as plain integers in inference events regardless of whether they came from a real tokenizer or from a heuristic approximation (wordCount / 0.75 for input, charCount / 4 for output).

For non-English text, code-heavy prompts, or models with non-GPT tokenizers (Gemma, LLaMA, Phi), the approximation error is frequently 2–3x. The backend and dashboard have no way to tell which numbers to trust.

Proposed fix

Add a companion boolean field to both TextInputMeta and GenerationOutputMeta:

• tokenCountEstimated: Boolean — true when the approximation heuristic was used, false when a real tokenizer was provided by the caller

Wire format addition (backwards-compatible — field absent on old clients):

"token_count": 42,
"token_count_estimated": true

This lets the backend suppress or annotate estimated values in analytics, and lets the dashboard surface a nudge to provide a real tokenizer.

No breaking changes required

The field is additive. Old clients omitting it can be treated as estimated by the backend.