arm=branch tag=b1 ready_s=238 at=2026-09-13T21:42:19Z
NOTE: transcribed from the session. b1 and m1 were run by hand, before
rig/boot_and_record.sh existed; b2/b3/m2 are that script's own output.
== 21:42:26 ==
source_digest: 9e68117a78a7480d
server_info:   {'max_context': 180736, 'max_batch': 8}
kv metrics:
arbi_serve_kv_pool_pages_total{backend="paged_kv:tkv-k4v4",kind="paged_kv"} 707.0
arbi_serve_kv_pool_pages_total{backend="",kind="gdn"} 707.0
--- boot lines ---
This was the only boot in the campaign with a COLD JIT/compile cache for its
tree: it built arbi_serve_exl3_a_prep_v3 from source (ptxas observed at 102%
CPU) and took 238 s against the 61 s a warm branch boot takes.
Its graph pool measured 1552.0 MiB. Boot b2 read that cached figure and
REPLACED it with its own 1440.0 MiB:
  graph_pool budget cache: REPLACING the cached 1627389952 B (1552.0 MiB)
  with this boot's 1509949440 B (1440.0 MiB) -- the reading is the cuMem
  per-tag counter (capture.cudagraphs + capture.io_buffers), which is exact
  for this cache key, so the monotonic ratchet does not apply.
112 MiB / 4.25 MiB per page = 26 pages, against the 23 pages b1 came short.
