Use a 5 s random-access measurement time limit

A 10 s control run on this branch moved by the same amount against develop
as the 5 s run, so halving the time limit does not change the results; it
only halves the benchmark step.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UeniB9mYWJs2z49XPgvbPi
Signed-off-by: Joe Isaacs <joe.isaacs@live.co.uk>

Author: joseph-isaacs

scripts/random-access-split.py

@@ -24,6 +24,8 @@
 LOCAL_ONLY_FORMATS = {"arrow-ipc", "litdata"}
 PATTERNS = ["correlated", "uniform"]
 OPEN_MODES = ["cached", "reopen"]
+# Seconds each measurement runs for, passed as --time-limit.
+TIME_LIMIT_SECS = 5
 
 
 def drop_os_caches() -> None:
@@ -98,6 +100,8 @@ def run_combinations(emit_ingest_records: bool, remote_data_dir: str | None) ->
                         pattern,
                         "--open-mode",
                         open_mode,
+                        "--time-limit",
+                        str(TIME_LIMIT_SECS),
                         "-d",
                         "gh-json",
                         "-o",