obsv: six per-database metric setters have no writer; add a sampler

[EnRaiha]

Version / build tested against

origin/main @ 1ff3551

Deployment mode

Origin, single node (local)

Engine(s) involved

observability

Summary

Six of the per-database metric setters in DatabaseMetricsRegistry
(nodedb/src/control/metrics/database.rs) have no caller anywhere in the
tree, so the families they feed publish constants:

record_qps                       callers_outside=1   (pgwire sql_exec only)
set_mirror_lag_ms                callers_outside=1   (mirror observer)
set_memory_bytes                 callers_outside=0
set_storage_bytes                callers_outside=0
set_connections                  callers_outside=0
set_bridge_queue_depth           callers_outside=0
set_wal_latency_p99              callers_outside=0
add_maintenance_cpu_secs         callers_outside=0

render_prometheus is reached from the /metrics route, so
nodedb_database_memory_bytes, nodedb_database_storage_bytes,
nodedb_database_connections, nodedb_database_bridge_queue_depth,
nodedb_database_wal_commit_latency_p99_us, and
nodedb_database_maintenance_cpu_seconds_total are exported with values that
never move. There is no sampler task that fills them.

I found this while wiring loader-facing counters: a bulk writer wants to read
per-database queue depth and commit latency to size its pacing, and the
gauges it would read are always zero. qps is also partial: only the pgwire
statement path calls record_qps, so HTTP and native queries do not count.

Steps to verify

python3 - <<'PY'
import pathlib, re, subprocess
src = pathlib.Path('nodedb/src/control/metrics/database.rs').read_text()
for m in re.findall(r'pub fn (\w+)', src):
    out = subprocess.run(['grep','-rn',f'.{m}(','nodedb/src','--include=*.rs'],
                         capture_output=True, text=True).stdout
    sites = [l for l in out.splitlines() if 'control/metrics/database.rs' not in l]
    print(f"{m:32} callers_outside={len(sites)}")
PY

Expected behavior

• A periodic sampler (order of ten seconds) fills each per-database gauge
  from its owning subsystem: connections from the connection tracker,
  memory and storage from the governor or catalog, bridge queue depth from
  the bridge dispatcher, WAL commit latency from the group-commit path.
• bridge_queue_depth becomes readable per database, so a bulk loader can
  back off before the WFQ refuses (the refusal now reports a retryable
  dispatch_capacity class).
• Where a source genuinely does not exist yet (maintenance CPU seconds is a
  candidate), the metric is either removed or documented as not yet filled,
  rather than exported as a constant.

Actual behavior

The setters exist and render; nothing calls them.

What actually happened? (check all that are true)

• Acknowledged/committed data was lost, corrupted, or silently wrong
• The server crashed, hung, or failed to start
• A security or isolation boundary was crossed
• Core functionality is broken with no acceptable workaround
• A workaround exists (dashboards must ignore the families; pacing stays
  conservative)

Proposed severity

SEV-4, observability. No correctness impact, but the containers that promise
backpressure visibility are wrong as published.

Reproducibility

Always; the audit above is deterministic.

Last known-good version / commit (if a regression)

Not a regression as far as I can find; the setters appear to have been added
without a sampler.

Environment & logs

Linux x86_64, single-node origin. Related: #371 added the retryable
dispatch_capacity class and its process-wide counter, and #372 wired the
graph write/delete counters; this issue covers the per-database family those
sit beside.

Before submitting

• I searched existing issues and this is not a duplicate.
• I reproduced this on current main (not a stale branch).
• This is not a security vulnerability.

[AubaidAhmedSaiyed]

Hello, I would like to work on this issue

[EnRaiha]

You may, just send your pr after you ready