docs(geode): SQLite knobs A/B verdict — no winner, knobs stay off by default

Backlog items 4-5 close. Both-variants-in-one-run A/B at 50k, repeated
with relay order reversed: every delta flipped with the order (the
second-running relay won queries and ingest latency in BOTH runs), so
readers=8 / mmap_size=256MiB / temp_store=MEMORY / periodic PRAGMA
optimize are all noise-level on container-class hardware. The config
plumbing stays (hardware-dependent, operators should measure their own
box); the example config now says so explicitly.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TtDNpayEYvJH7QuPswND3A

PLANS.md

@@ -38,7 +38,7 @@ listing (including archived plans), open that folder's `README.md`.
 | cli | 6 | 5 | 1 | 0 | 0 | [cli/plans](cli/plans/README.md) |
 | quic | 4 | 3 | 0 | 0 | 1 | [quic/plans](quic/plans/README.md) |
 | quic/interop | 1 | 1 | 0 | 0 | 0 | [quic/interop/plans](quic/interop/plans/README.md) |
-| geode | 4 | 4 | 0 | 0 | 0 | [geode/plans](geode/plans/README.md) |
+| geode | 5 | 4 | 0 | 0 | 0 | [geode/plans](geode/plans/README.md) |
 | docs (frozen) | 52 | 48 | 2 | 0 | 2 | [docs/plans](docs/plans/README.md) |
 
 ## Live work (not shipped)

geode/config.example.toml

@@ -46,6 +46,11 @@ path = "/"
 in_memory = false
 file = "/var/lib/geode/events.db"
 
+# NOTE: the four knobs below measured as pure noise in relayBench A/B
+# runs on 4-core container hardware at 50k events (see
+# geode/plans/2026-07-04-sqlite-knobs-ab.md) — their value is
+# hardware-dependent, so benchmark on YOUR box before enabling.
+#
 # Reader-connection pool size (file-backed stores only). Default 4.
 # readers = 4
 

geode/plans/2026-07-04-sqlite-knobs-ab.md

@@ -0,0 +1,50 @@
+# Server-side SQLite knobs — A/B verdict: no winner on container hardware
+
+**Status: closed (negative result recorded).** Backlog items 4–5 of the
+relay performance campaign.
+
+## What was tested
+
+The knobs the campaign had previously reverted as noise-inconclusive,
+re-tested under the A/B protocol — both variants in ONE relayBench run
+(identical container conditions), then the run repeated with the relay
+order reversed to expose order bias. 50k synthetic corpus, all four
+knobs on the candidate at once:
+
+```toml
+[database]
+readers = 8                     # reader pool (default 4)
+mmap_size = 268435456           # 256 MiB
+temp_store_memory = true
+optimize_interval_seconds = 15  # PRAGMA optimize between ingest and queries
+```
+
+## Result
+
+Every observed delta flipped with run order. In BOTH runs the
+second-running relay won most query scenarios and ingest latency —
+regardless of which variant it was:
+
+- run 1 (plain → knobs): knobs won 7/10 query scenarios, better ingest
+  p99 (8.2 vs 16.0 ms).
+- run 2 (knobs → plain): plain won 6/10 query scenarios, better ingest
+  p50/p99/throughput (8,552 vs 8,015 ev/s).
+
+Storage size identical (±1 MiB). The periodic `PRAGMA optimize` also
+did not move the author-archive/planner-sensitive scenarios.
+
+## Verdict
+
+**All knobs stay off by default.** The `[database]` config plumbing
+ships anyway (readers / mmap_size / temp_store_memory /
+optimize_interval_seconds) because their value is hardware-dependent —
+an operator on NVMe with a large page cache or a memory-constrained VPS
+should measure on their own box — and `optimize_interval_seconds`
+remains sensible operationally for long-running relays even without a
+measurable win at 50k-fresh-database scale.
+
+**Do not re-benchmark these on container-class 4-core hardware** without
+first fixing the run-order bias: the protocol note in
+`quartz/plans/2026-07-03-incremental-negentropy-storage.md` applies —
+same-run A/B plus order reversal is the minimum, and anything that
+doesn't survive the order flip is noise.

geode/plans/README.md

@@ -1,6 +1,6 @@
 # geode plans
 
-_Audited 2026-06-30. 4 plans: 4 shipped (archived), 0 in-progress, 0 queued, 0 abandoned._
+_Audited 2026-06-30. 5 plans: 4 shipped (archived), 0 in-progress, 0 queued, 1 closed (negative result)._
 
 Performance-focused design docs for future work. Each file is a
 self-contained sketch — problem statement, observed numbers, proposed
@@ -26,3 +26,9 @@ regressions show up in the regular CI matrix once they're enabled.
 - [archive/2026-05-07-live-broadcast-fanout-index.md](archive/2026-05-07-live-broadcast-fanout-index.md)
 - [archive/2026-05-07-connection-scaling.md](archive/2026-05-07-connection-scaling.md)
 - [archive/2026-05-07-negentropy-large-corpus.md](archive/2026-05-07-negentropy-large-corpus.md)
+
+## Closed (negative result)
+
+| Plan | Verdict |
+| ---- | ------- |
+| [2026-07-04-sqlite-knobs-ab.md](2026-07-04-sqlite-knobs-ab.md) | readers/mmap/temp_store/PRAGMA-optimize knobs: no winner on container hardware; every delta flipped with run order. Knobs ship config-gated, off by default. |