fix(observability): stop single-binary Tempo evicting its only ingester (closes #156) (#157)

## What & why

`verify-tracing` flaked on `verify-stack` run 722 — `FAIL — no single trace spanned ['bff', 'projection-api']` — and went green on a plain re-run of the same commit. **The trace chain was not broken; Tempo could not ingest:**

```
removing distributor_pool failing healthcheck addr=127.0.0.1:9095
  reason="rpc error: code = DeadlineExceeded"
pusher failed to consume trace data  err="context canceled"   (x18)
```

The root cause is the *mechanism* of the data loss, not whatever caused the stall. Tempo runs **single-binary**, so the distributor and the ingester are the same process and the distributor's ingester pool holds exactly one, in-process, member. dskit nevertheless health-checks that member over loopback gRPC with a **1 s** deadline (`checkinterval: 15s`, confirmed from the running image's `/status/config`). On the shared runner a transient stall blows the deadline, the only ingester is evicted from the pool, and every subsequent push fails until the next check interval — spans silently dropped.

With one in-process ingester the health check can **never** route around a failure. Its only possible effect is to discard data. So it is off:

```yaml
ingester_client:
  pool_config:
    healthcheckenabled: false
```

This lands at the point where *both* candidate triggers named in #156 (GC pressure near `mem_limit`, CPU contention from the grown stack) turn into lost spans, so **`mem_limit: 400m` is untouched** — raising it on a memory-tight runner risks reintroducing the `verify-e2e` OOM of #144. It also does not paper over anything the way a longer `TRACING_TIMEOUT` would (#156's own note).

Second change: `infra/tracing-check.py` prints `tempo_distributor_ingester_clients` on its failure path. From the check's side, Tempo-dropped-spans and missing instrumentation look identical — that ambiguity is what cost a container-log dive on run 722. A recurrence now names itself.

Closes #156

## Definition of Done

- [x] Linked Gitea issue (#156).
- [ ] **Failing test committed before the implementation — N/A, and deliberately so.** The trigger is runner load, so no deterministic red exists; the "red" is run 722's observed `verify-tracing` failure plus its Tempo logs. Same precedent as d5e5fa2 (#115, Playwright OOM) and 4aafd32 (#147, uWSGI caps). A test asserting the config says what the config says would add no gate: Tempo hard-fails on an unknown key (verified — `field health_check_enabled not found in type client.PoolConfig`), so a typo or a config rename on a Tempo bump already turns `verify-up` red.
- [x] Conventional Commits referencing the issue (`refs #156`).
- [ ] CI green — the point of the change.
- [x] `docker compose up` health unaffected (Tempo is not in `WAIT_SVCS`; config-only change, same image).
- [x] Docs updated — ADR-0023 Consequences.
- [x] ADR — amended **ADR-0023** rather than adding a new one: this is a consequence of that ADR's single-binary Tempo choice, not a new decision (one decision per ADR, §12).
- [x] Demo note — N/A, not user-visible.

## Notes for reviewers

Verified locally against the built image (the flake itself is not locally reproducible — see the runner-load point above):

1. `docker run --rm register-referentie/tempo:dev -config.file=/etc/tempo.yaml -config.verify=true` → parses.
2. `GET /status/config` on the running container → `healthcheckenabled: false` (was `true`).
3. The new diagnostic reads `tempo_distributor_ingester_clients` off a live Tempo.

Worth knowing: that metric is legitimately `0` on an idle Tempo — the pool is populated lazily on first push. It only prints on the failure path of a check that has already generated traffic, so the reading is meaningful there, but don't read a bare `0` on a quiet stack as an eviction.

Follow-up left undone: if `verify-tracing` still flakes after this, the next suspect is the .NET OTLP exporter timeout (#156's last note), not Tempo's memory cap.Reviewed-on: #157

docs/architecture/adr-0023-observability-stack.md

@@ -67,6 +67,14 @@ itself, so no in-image healthcheck tool is required.
 - Three more images built each CI run (kept small; not on the health-gate list).
 - Storage is ephemeral container fs — a demo backplane, not a retention target.
   Object storage for Tempo / remote-write for Prometheus is a later concern.
+- Tempo runs **single-binary**, so its distributor and ingester are one process and
+  some of its distributed-mode machinery is not just redundant but harmful. Its
+  ingester-pool health check is disabled (`ingester_client.pool_config`) because with
+  a single in-process ingester the check can never route around a failure — a 1s
+  loopback-gRPC deadline missed under CI load only evicted the one ingester and made
+  Tempo drop spans, which is how `verify-tracing` flaked (#156). Expect the same
+  shape from other distributed-mode knobs if we tune them; the fix is to switch to
+  real multi-ingester Tempo, not to re-enable them here.
 
 ## Coupling rules touched (CLAUDE.md §8)