Traces and metrics — the deferred half of @btravstack/observability

[btravers]

From a gap analysis against NestJS and AdonisJS.

packages/observability/CLAUDE.md and the root CLAUDE.md's Deferred, deliberately section already describe this; opening it as an issue so it is trackable rather than only narrated.

The package is named for the whole because logs, traces and metrics share a correlation id, a resource, a config slice and a flush-on-shutdown lifecycle. Only the logging half ships. The root CLAUDE.md is explicit that the rest must never be described as shipped.

The shape is already written down in packages/observability/CLAUDE.md: Tracer / Meter ports, the OTel NodeSDK as a resourceful provider whose release flushes, a span per unit through StartOptions.unit, and W3C traceparent feeding UnitMeta.traceId in the three transport starters — which is the piece that makes a trace started elsewhere join a trace here, and is why traceId is a runtime-suppliable field in the first place.

Two things make this more than a port to write:

• The flush-on-shutdown lifecycle is the kernel's problem, and the kernel already solves it. A resourceful provider's release runs on every exit path including startup failure, and the drain's three beats give a real window to flush in. Getting spans out before the process dies is the part most OTel integrations get wrong; here the primitive exists.
• The auto-instrumentation constraint noted in the package's own CLAUDE.md does not go away, and should be settled before the port is written rather than after.

Acceptance

• Tracer and Meter ports with an OTel-backed provider, flushing through release.
• A span per kernel unit, correlated with the same traceId the logger already stamps.
• traceparent honoured inbound by @btravstack/http, @btravstack/amqp and @btravstack/temporal, feeding UnitMeta.traceId.
• The Deferred, deliberately entries in the root and package CLAUDE.md removed when it lands — and until then, still never described as shipped.