feat(tracing): add OTel observability layer for TAC voice calls

[bsahajsinghani]

Summary

• Moves OTel packages to an optional tac[tracing] extra — no longer a hard dependency for all TAC users
• Auto-initializes tracing in TAC.__init__ when OTEL_ENABLED=true — zero manual setup required by the developer
• Changes default OTEL_ENDPOINT from a Twilio-internal URL to http://localhost:4318 (standard OTel collector port, works out of the box with Jaeger, Grafana Tempo, Honeycomb, Datadog)
• Adds graceful ImportError handling in setup_tracing() with an actionable warning when packages are not installed
• Adds tests/test_tracing.py: 19 tests covering no-op behaviour, start/end call lifecycle, all span types, turn counter, and traceparent injection

Span hierarchy for a voice call

call  [root — full call duration]
├── call.co_init              CO polling latency at call start
├── call.profile_lookup       phone → profile_id resolution
├── call.first_prompt_wait    how long first prompt waited for init_task
└── turn  [one per user utterance]
    ├── memory.recall
    │   ├── memory.profile_lookup
    │   ├── memory.profile_fetch
    │   └── memory.recall_api    (vector search only)
    └── llm.completion
        ├── llm.prompt_build
        └── llm.response_stream
            └── event: first_token_sent   (LLM TTFT marker)

Why this matters — latency data

Instrumented via this tracing layer across 4 experiments on a SageMaker-hosted Qwen2.5-1.5B voice agent:

Experiment	TTFA	app_lat	co_init	profile_lookup	first_prompt_wait
Baseline	1322ms	1104ms	940ms	753ms	—
Early co_init (9h)	784ms	549ms	3607ms	528ms	—
Parallel init (9i)	1054ms	842ms	3316ms	390ms	—
Pre-fetch (9j)	1274ms	1062ms	3479ms	467ms	0ms

first_prompt_wait=0ms across all 9j runs confirms that pre-fetching observations inside the init task is safe — CO poll always completes before the user finishes speaking.

Usage

pip install tac[tracing]

OTEL_ENABLED=true
OTEL_ENDPOINT=http://localhost:4318   # or your Honeycomb/Datadog/Grafana endpoint
OTEL_SERVICE_NAME=my-voice-agent

That's it — no code changes needed. Tracing starts automatically when TAC is instantiated.

Test plan

• uv run pytest tests/test_tracing.py — 19 tests pass
• uv run pytest — 711 tests pass, no regressions