feat: adopt current OTEL cached token conventions

[bml1g12]

Context

Today, when we emit the gen_ai.usage.input_tokens metric, we calculate this already correctly in LiveKit (as defined by https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/) however two gaps exist in LiveKit today:

• The cached text token count is emited to a custom namespace not recognised by OTEL or major providers like LangFuse or Pydantic (see "## Removed trace type constants (not in the OTEL spec)" section below)
• Audio token counting is not part of official OTEL syntax, but recently LangFuse fix(otel): normalize gen ai usage details langfuse/langfuse#13110 and pydantic More detailed token usage span attributes and metrics open-telemetry/semantic-conventions#1959 (comment) have adopted a convention for this that would be useful LiveKit to follow, as it allows cost estimation of gpt-realtime in LangFuse. LangFuse has recently merged a fix on their side to correctly map values stored in the convention used in this PR to their internal cost modelling .

Summary

Aligns realtime and related LLM spans with current OpenTelemetry GenAI usage conventions and Langfuse OTEL ingestion (langfuse#13110): top-level gen_ai.usage.input_tokens / output_tokens represent text counts; cache read uses the standard dotted key; audio breakdowns use gen_ai.usage.details.* so backends can normalize usage without double-counting cache or mixing modalities.

Internal metrics (RealtimeModelMetrics, ModelUsageCollector, Prometheus counters) are unchanged—only trace attributes and trace_types constants are updated.

Attribute mapping

Realtime spans (record_realtime_metrics) now set:

Intent	OTEL / Langfuse-oriented attribute	Source on RealtimeModelMetrics
Top-level input (text-only, inclusive of cached text)	gen_ai.usage.input_tokens	input_token_details.text_tokens
Top-level output (text-only)	gen_ai.usage.output_tokens	output_token_details.text_tokens
Cached text (cache read)	gen_ai.usage.cache_read.input_tokens	cached_tokens_details.text_tokens (if non-zero)
Input audio (non-cached)	gen_ai.usage.details.input_audio_tokens	input_token_details.audio_tokens
Cached audio read	gen_ai.usage.details.cache_audio_read_tokens	cached_tokens_details.audio_tokens
Output audio	gen_ai.usage.details.output_audio_tokens	output_token_details.audio_tokens

gen_ai.usage.cache_creation.input_tokens is not set until a provider exposes cache-creation counts in our metrics model.

Removed trace type constants (not in the OTEL spec)

These ATTR_GEN_AI_USAGE_* aliases were removed because their string keys were informal (LiveKit/Langfuse-style flat names) and not defined in OpenTelemetry GenAI semantic conventions:

Removed constant	Removed attribute key
ATTR_GEN_AI_USAGE_INPUT_TEXT_TOKENS	gen_ai.usage.input_text_tokens
ATTR_GEN_AI_USAGE_INPUT_AUDIO_TOKENS	gen_ai.usage.input_audio_tokens
ATTR_GEN_AI_USAGE_INPUT_CACHED_TOKENS	gen_ai.usage.input_cached_tokens
ATTR_GEN_AI_USAGE_OUTPUT_TEXT_TOKENS	gen_ai.usage.output_text_tokens
ATTR_GEN_AI_USAGE_OUTPUT_AUDIO_TOKENS	gen_ai.usage.output_audio_tokens

Top-level gen_ai.usage.input_tokens and gen_ai.usage.output_tokens remain; they now map to text token fields only on realtime spans, consistent with semconv + Langfuse normalization.

Other call sites

• Cascaded LLM spans (llm.py): set gen_ai.usage.cache_read.input_tokens when prompt_cached_tokens > 0.
• Eval judge spans (run_result.py): same cache-read key; redundant duplicate text keys removed.

Tests

• Unit tests assert the attribute dict produced by record_realtime_metrics (including omission of zero breakdowns).

---

Breaking note: External code that imported the five removed ATTR_GEN_AI_USAGE_* names must switch to the new constants (ATTR_GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS, ATTR_GEN_AI_USAGE_DETAILS_*, etc.).