otlp export + obs.init for observability

[kacy]

summary

phase 3 of the observability sprint: OTLP export and a one-call obs.init.

• std.otlp encodes std.trace spans and std.metrics samples as OTLP
  protobuf (ExportTraceServiceRequest / ExportMetricsServiceRequest),
  hand-built on std.protobuf the way the db drivers hand-write their wire
  protocols, and POSTs them to an OTLP/HTTP collector at /v1/traces and
  /v1/metrics with content-type: application/x-protobuf.
• std.obs init() reads the standard OTEL_* environment variables,
  turns tracing and metric collection on, and spawns a background exporter
  that periodically flushes finished spans and the current metric snapshot.
  with no OTEL_EXPORTER_OTLP_ENDPOINT set it does nothing, so the call is
  free to leave in.

public api:

• otlp.target(service_name, endpoint, protocol) -> Target
• otlp.encode_traces(service_name, spans) -> Bytes
• otlp.encode_metrics(service_name, samples) -> Bytes
• otlp.export_traces(target, spans) -> Bool / otlp.export_metrics(target, samples) -> Bool
• obs.init(), obs.start(service_name, endpoint, protocol, metric_interval_ms, span_delay_ms), obs.active(), obs.shutdown()

scope

spans with string attributes and status, counter and gauge metrics in full,
histograms as a count+sum data point (per-bucket boundaries omitted — noted in
the code), OTLP/HTTP only (grpc fails with a clear message).

tested

• pith test std/otlp.pith (6 tests): endpoint parsing plus reader round-trips
  of a trace request, a monotonic-sum counter with an int data point and its
  label attribute, and a histogram count+sum.
• pith test std/obs.pith (3 tests): stays off with no endpoint, loop tick
  picks the smaller interval, empty endpoint is a no-op.
• doc --check clean on both modules.
• end-to-end: ran obs.init against a local stub collector that decodes the
  posted protobuf; it received /v1/traces (service.name + span name) and
  /v1/metrics (service.name + metric name) as expected.
• make run-examples: 86 passed, 0 failed.

design notes

otlp.Target is named Target rather than Config to avoid a cross-module
struct-name clash with tls.Config (E219). histograms map to a
count+sum data point for v1; a fuller mapping with bucket boundaries can follow.