blog: SQLite vs DuckDB benchmark + ingest admission gate

[jstojiljkovic]

Benchmark series post 3: SQLite vs DuckDB telemetry backends on the same Hetzner CCX13.

• website: new post sqlite-vs-duckdb.mdx plus read-probe charts and dashboard screenshots. Every number in the post is backed by the raw JSONs in this PR.
• backend: ingest admission gate (INGEST_MAX_CONCURRENT, INGEST_ADMISSION_WAIT_SECONDS) that caps concurrent ingest processing and sheds overload with 503 + Retry-After instead of getting OOM-killed.
  Rejects are counted and exposed as ingestRejected in /api/health/deep.
• loadgen: post-fill digestion gate (--max-digest-wait) so read probes measure steady-state query cost rather than a mid-WAL-checkpoint engine; fills now count only 2xx-confirmed rows so admission 503s
  don't produce phantom rows; workflow dispatch overrides for the read threshold and the Phase 2 rate ladder.
• data: throughput, read-probe, and 60-second diagnostic JSONs per signal in benchmarks/blog/post-3-data/.