Instrumentation: enterWith leak, operation span reachable from the public API, context-store eviction, body-drain diagnostics, DEXPACE_LOG_LEVEL wiring #80

Part of the audit remediation umbrella #67. Milestone 4. Severity: HIGH (first two items), MEDIUM (rest). Requirements: OBS-20, OBS-22, OBS-23, OBS-29, OBS-35, CTX-11, CTX-14, CTX-15, CTX-16.

Current SDK behavior

• AsyncLocalStorage.enterWith leaks outward. pushDiagnosticFields
  (observability/diagnostic-context.ts:91) and createAsyncScopedStore.enter (:118,125)
  call enterWith. The restore closure runs after awaits inside send(), in a different
  async resource, and never reaches the caller. After runtime.send() returns,
  getActiveSpan() is the ended span and the process-wide diagnostic context still holds that
  request's trace.id / span.id. An unrelated application log through any core Logger then
  carries them. The existing tests assert inside the same continuation that called close().
• The operation span is suppressed after the first send. startOperationSpan
  (pipeline/runtime.ts:50) bails when getActiveSpan().isRecording. The leaked span still
  reports isRecording === true, so only the first operation per process gets the
  http.client.operation span.
• The operation span cannot be enabled at all from the public API. PipelineBuilder.build()
  (pipeline/builder.ts:241) calls createRuntime(flattened, transport) with no
  contextInit; createRuntime is @internal; standardResilience() has no instrumentation
  option. Every public runtime gets noopInstrumentationBundle. createInstrumentationBundle
  is public with nowhere public to pass its result. The changeset
  2026-09-04-per-operation-span.md shows createRuntime(...) in its example.
• A throwing tracerFactory leaks the context-store entry. runtime.ts:170 installs the
  request context before the try at :176; startOperationSpan and activateSpan (:174-175)
  sit in the gap. store.size grows by one per failed send.
• span.end() can run twice. If end() on the success path throws, the catch calls
  recordException and end() again (runtime.ts:193,198-199).
• Body-drain failures are silent. The catches at observability/logging-step.ts:309-312
  and :333-335 return an empty capture and emit nothing. A fileBody() over a deleted file
  logs "http.request.body.preview": "" with no http.instrumentation.* event.
• DEXPACE_LOG_LEVEL is dead by default. resolveGranularity (logging-step.ts:84)
  reads getGlobalConfiguration(), whose default is an empty configuration with an env seam of
  () => undefined. Nothing in packages/ or docs/sdk-documentation/ installs
  defaultConfiguration(). The key name is hardcoded (:85); OBS-35 says the SDK MUST NOT bake
  in a default key name.
• Smaller: noopInstrumentationBundle.activeSpan is undefined instead of NOOP_SPAN
  (context/instrumentation.ts:65). operationName is never passed to promoteToRequest
  (runtime.ts:169), so ctx.context.operationName is always undefined (CTX-16).
  redactUrl(string) normalizes port, path and host case (redaction.ts:83; OBS-14).
  logging-step.ts:427's span is ended inside executePipeline, not in the finally.

Expected behavior

When send() resolves or rejects, the active span and the diagnostic fields are what they
were before the call. Every operation gets its own http.client.operation span. A consumer
can pass an InstrumentationBundle through PipelineBuilder and standardResilience().
The context store has the same size after a failed send as before it. A body-drain failure
emits a best-effort http.instrumentation.* diagnostic. DEXPACE_LOG_LEVEL (or a
caller-chosen key) is read from the environment without extra wiring, or the required wiring
is documented.

Notes and leads

• Replace enterWith plus restore closure with storage.run(value, fn) around the awaited
  body of send(). run restores on exit and does not leak outward. Where a handle-based API
  must stay, document that close() cannot restore the caller's context. Test: log after
  await send() and assert no trace.id; assert the second send gets its own operation span.
• Public path: add an instrumentation (or contextInit) option to PipelineBuilder's
  constructor and to standardResilience()'s options, threaded to createRuntime. Pass
  operationName through the same option (CTX-16). Fix the changeset example.
• Move install inside the try, or open the try before the span start. Single exit for
  span.end() guarded by an ended flag.
• Body-drain: safeEmit an http.instrumentation.bodyCaptureFailed event with cause().
• Log level: either default the global configuration to defaultConfiguration() (behavior
  change, changeset) or document setGlobalConfiguration(defaultConfiguration()) in
  pipelines.md. Add configKey to LoggingStepSettings.
• Land subtask 1 first: tracerFactory's TSDoc and the deviations.md OBS-29 row change again
  here.
• Minor changeset for @dexpace/core (new builder option). api:local.