codex exec does not export the `codex.turn.token_usage` OTLP metric (0.144.x), despite #12913 being closed

Summary

With OpenTelemetry configured via [otel] exporter = { otlp-http = {...} }, codex exec (the non-interactive CLI) exports traces and log events to the configured OTLP collector but exports no OTLP metrics - in particular the documented codex.turn.token_usage histogram is never emitted. The interactive TUI exports it. As a result, per-run token/cost usage is invisible to OTLP metric-based monitoring for any automated / CI / agent workflow that drives codex exec.

Notably, the per-turn token counts are present in the exported OTLP - but only as span/log attributes, not as a metric instrument (details below).

Environment

• codex-cli 0.144.3 (reproduced current; 0.144.4 has no user-facing changes and 0.144.5 is a dangerous-command-detection fix only - no telemetry changes, so this applies to latest)
• Linux; OTLP export via otlp-http to a local collector

Repro

1. ~/.codex/config.toml:

``toml
[otel]
environment = "dev"
exporter = { otlp-http = { endpoint = "http://127.0.0.1:4319", protocol = "binary" } }
``

1. Run a minimal OTLP sink on 127.0.0.1:4319 that records incoming POST bodies.
2. codex exec --sandbox read-only "print OK"
3. Inspect the captured OTLP payloads (5 POSTs, ~11 KB).

Observed

• OTLP POSTs arrive carrying codex traces / log events: codex.startup_phase, codex.api_request, codex.sse_event, codex.websocket_request/connect, codex.user_prompt, codex.turn_ttft, codex.conversation_starts.
• Per-turn token counts are present only as span/log attributes: input_token_count, output_token_count, cached_token_count, reasoning_token_count, tool_token_count.
• The OTLP metric codex.turn.token_usage (histogram, with token_type) is absent - no metric datapoints are exported under codex exec at all.
• The CLI prints tokens used N to stdout, so the totals exist internally.

Expected

codex exec exports the codex.turn.token_usage metric (histogram, by token_type: input / cached_input / output / reasoning_output) to the configured OTLP collector, the same as the interactive TUI - so headless usage can be monitored via OTLP metrics.

Not a duplicate

• #12913 reported this and was closed via #13083 ("Enable analytics in codex exec and codex mcp-server"), which defaulted analytics-on at exec OTel init. Per the repro above, the OTLP token_usage metric is still not exported under codex exec.
• Distinct from #18600 (structured session metrics in exec stdout output, not OTLP) and #27467-#27469 (exec-server metrics, not the one-shot codex exec CLI).

Impact

Headless / automated codex exec usage (CI pipelines, multi-agent orchestration) cannot surface token/cost usage into OTLP metric dashboards, forcing out-of-band parsing of ~/.codex/sessions/**/rollout-*.jsonl (or scraping the log attributes) to recover per-run token counts. Emitting the already-computed counts as the documented metric would close the gap.