bench(db): record the cluster suite on v0.8.7 — the per-vhost forward hop

[luthermonson]

The cluster suite has been authored and gate-validated since #4 but never recorded — lanes P2 and P3 need per-site clustered replication (ephpm#416), which was in no published image. The v0.8.6/v0.8.7 images were published 2026-09-01, so all five lanes now run in their intended mode for the first time.

Recorded on ephpm/ephpm:v0.8.7-php8.5 (image 85c3444cedcd), which also carries ephpm#429 — replicate per-site databases on bridge-only nodes.

Run health

All gates passed, 52/52 cells 100% HTTP 200, suite exit 0. The mode gate confirmed per-site clustered mode on all three nodes; ownership resolved to exactly one node (cdc_subscribers=2 on n2, elected as SQLite primary on n2 alone); per-site replication converged owner → non-owner before any cell was measured.

The three deltas

Comparison	Result	Verdict
S → W — cost of clustering	−11.8% read / −13.0% write at c=1	Unresolved under this file's 20% rule (4–7% rep spread); bounded above at ~13%. Not the 0 RPS write collapse the historical sqld lane recorded.
P1 → P2 — cost to a tenant	−16.7% (c=1), −28.1% (c=16), write path	Resolved. Both c=1 cells have 0.2–0.3% rep spread, so the effect is ~50× the noise. wire-point control flat.
P2 → P3 — the sql/<site> hop	−69.4% read / −53.7% write at c=1	Strongly resolved — 0.0–3.8% spread against 54–69% effects.

The headline number decomposes cleanly. The two fixtures agree independently on a per-statement cost:

• read, 10 forwarded statements: p50 1.181 → 3.929 ms ⇒ 275 µs/statement
• write, 1 forwarded statement: p50 1.299 → 1.554 ms ⇒ 255 µs/statement

Within 8% of each other across a 10× difference in statement count — the hop is paid per statement, not per request, so a ten-query page on a non-owner pays it ten times.

The wire-point control did its job. On the non-owner the non-forwarded path is not slower (+6.1% at c=1, +1.2% at c=16, sub-1% spread). Had it also been slow, the slowdown would not have been forwarding and the whole P2→P3 story would have collapsed into "n1 is slower than n2". It didn't.

Flagged rather than smoothed over

• P1 bridge-point is anomalous — 494/553 RPS, 11.3% spread, p99 6–9 ms, and 36% slower than the write cell on the same node, which is true in no other lane. Taking it at face value would claim per-site clustering makes reads 58% faster. The read-path P1 → P2 delta is therefore reported as not measurable from this run, pending a re-record of P1 alone.
• P3 bridge-write has a 10× p50→p95 tail at c=1 (1.55 ms p50, 15.4 ms p95, 25–26 ms p99) where the owner's cell sits at 2.3 ms p95. Read path shows no such tail.
• The divergence probe reproduced exactly — a stock pdo_mysql write on the non-owner left the owner at 91463 and the writer at 91464. That fix (ephpm#432) is not in v0.8.7, so these numbers document the asymmetry as it ships. Per-site clustered mode remains experimental (Turso is Beta upstream).

Contents

• docs/cluster-persite-v087.md — full write-up, both reps per cell, per-statement cost model, observations for ePHPm
• DB-BENCH.md — Reference Numbers section filled in (was "None yet")
• Default cluster image bumped v0.8.5 → v0.8.7 in scripts/run-db-bench.sh and db/bench-cluster.sh, and the "v0.8.6 images are not on Docker Hub" notes dropped — no longer true

Raw oha output is gitignored per repo convention and kept locally.