- Move src/telemetry/export/writers.ts to writers/json.ts and the
  matching test to writers/json.test.ts. The writer keeps the same
  shape; the doc comment is tightened and the test drops a
  redundant atomic-failure assertion already covered by
  writeAtomically's own tests.
- Extract parseTelemetryTimestampMs from range.ts's
  isTimestampInRange so other export writers can reuse it.
- Add a trivial asyncIterable test helper to test/mocks/.
- Inline the duplicated `vi.mock("node:fs", ...)` /
  `vi.mock("node:fs/promises", ...)` factory pair (9 lines per file)
  as a 2-line dynamic import in the affected test files.

Upstream-compatible groundwork for the OTLP writer branch (#961);
isolates move + helper-extraction churn so that PR's diff stays
focused on OTLP-specific work.

- Add a required onCleanupError callback to writeAtomically and
  writeJsonArrayExport so callers cannot silently lose temp-file
  cleanup diagnostics (matters on Windows, where AV/Indexer locks
  can leave large export orphans). cliCredentialManager passes its
  logger to restore the warn-on-cleanup behavior that existed
  before the extraction.
- Wrap the rm+callback chain in a try/catch so a throwing
  onCleanupError can't replace the original write error.
- Document that writeAtomically requires the parent directory to
  exist.
- Use suffix "temp" so the call matches the doc example and the
  cliManager convention.
- Cover the cleanup-failure path with new tests: callback invoked
  with the rm error, and a throwing callback still surfaces the
  writer's error.
- Tighten the json writer's midstream-error test to assert the
  destination survives and no temp file is left behind.
- Fix the asyncIterable helper's docstring: it exercises async
  iteration, not stream backpressure.

After the 999->888 takeover, the 888 monitor would eventually fail
network reads, call processLost, then searchForProcess would find
888 again and emit ssh.process.recovered (same-PID rediscovery).
On slow CI runners that race won the negative assertion at
sshProcess.test.ts:469. Return [] from find after the takeover so
no rediscovery is possible.

Wire-format fixes:
- Switch http.requests count_* sums from DELTA to CUMULATIVE temporality
  (Prom/Mimir/Grafana Cloud reject the delta+sum combination).
- Maintain per-series running totals across the export, anchored at the
  first observed window's startTimeUnixNano.
- Suppress zero-cumulative counters so routes that never errored don't
  ship empty data points.

Semantic-convention compliance:
- Use OTel-standard `service.instance.id` (was `coder.session.id`) and
  `host.id` (was `coder.machine.id`) for portability with OTel-aware
  backends. The previous vendor keys were pure aliases.
- Add `schemaUrl` to each ResourceLogs/ResourceSpans/ResourceMetrics
  container and each scope container.
- Stamp scope.version with the extension version.