test(geode): keep deletion-settle benchmark as robust shape guard

Drop the flaky publish-into-large-relay warmup from DeletionSettleBenchmark
(it timed out the measured reconcile at N=100k — the container noise the
docstring already warns against) and remove the throwaway ScratchSettleTiming
investigation tool. Record in the docstring what the phase breakdown proved:
the settle's extra time over a bare reconcile is O(K) relay-ingest of the K
residual deletions, dominated by one-time JVM/JIT warmup of the publish path
(consecutive K-note batches fell ~3100->570ms), not the deletion algorithm and
not O(N).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JgL1WTV4Hkp2uuXcUHCHGt

geode/src/test/kotlin/com/vitorpamplona/geode/DeletionSettleBenchmark.kt

@@ -52,6 +52,14 @@ import kotlin.test.assertEquals
  * (the O(N) part it shares with any sync) next to the settle cost, so the deletion
  * overhead is visible as "≈ a couple of reconciles + K", not "+ a content re-download".
  *
+ * Why the printed settle can read as several× a bare reconcile at large N: the extra time
+ * is NOT the deletion algorithm (a phase breakdown showed reconciles stay ~sub-second at
+ * N=100k, and the settle re-fetches K=20, not N). It is entirely the K `publishAndConfirm`
+ * ingests into a large geode relay — publishing K *plain* notes costs the same — and that
+ * ingest path is JVM-cold on first use: consecutive K-note batches dropped monotonically
+ * (~3100 → ~570 ms) purely from JIT warmup. So the cost is O(K) relay-ingest dominated by
+ * one-time warmup, independent of N.
+ *
  * Default N is small so it doubles as a fast correctness guard; scale it with
  * `-DdelBenchN=200000` to see the shape at size. Not a speed assertion (container noise).
  */