BOARD B - SOAK AND COVERAGE

Two instruments run continuously on this board. Both are attached as sources.

=== 1. RESIDENT CANARY (768 MiB, mlocked, hash-verified, single-threaded) ===
first record         : 2026-09-14T05:29:07Z
last record          : 2026-09-17T16:30:16Z
service starts       : 8
passes completed     : 1277
faults captured      : 17

cumulative distinct frames watched (last coverage record):
  "rotation":656
  "vmlck_mib":768
  "cum_frames":1949752
  "cum_cov_pct":96.092
  "total_frames":2029056

  vmlck_mib is asserted every rotation. memtester-class tools silently SHRINK
  their buffer when mlock fails, which turns a 5 GiB test into a 200 MiB one
  that reports success. We assert the locked size rather than trusting it.

=== 2. NIGHTLY WIDE SWEEP (5,632 MiB in one window) ===
  2026-09-14T22:25:51Z	REFUSED	the board is claimed - claim : HELD for 0h 0m
  2026-09-14T22:26:07Z	2026-09-14	INCOMPLETE	no result record (rc=124) - see /srv/data/badram/sweep-2026-09-14.log
  2026-09-14T22:39:08Z	2026-09-14	min=13	passes=3	rotations=3	coverage=95.750%	covered_mib=7589	hits=0	rc=0
  2026-09-15T21:40:12Z	2026-09-15	min=61	passes=14	rotations=14	coverage=95.951%	covered_mib=7605	hits=0	rc=0
  2026-09-16T21:45:54Z	REFUSED	the board is claimed - claim : HELD for 1h 3m

  Reaching ~96% of RAM in a single window, hits=0 on both completed runs.
  REFUSED lines are the guard working as designed: the sweep will not run while
  the board is claimed for other work, and it records WHY rather than going
  silent - a missing sweep and a sweep that found nothing must not look alike.
  The INCOMPLETE line is our own bug, kept here deliberately: the backstop timeout
  was shorter than one pass and killed a healthy run before it could report.

=== WHY THE TWO INSTRUMENTS APPEAR TO DISAGREE, AND DO NOT ===
  The wide sweep covers ~96% of RAM and reports hits=0. The canary covers 768 MiB
  at a time and reports 17. That is not a contradiction - it is exposure time.

  canary : 1277 passes, continuous, ~3.5 days resident  -> 2 fault events
  sweep  : 2 completed runs of ~13 and ~61 minutes         -> 0 fault events

  At an observed rate of roughly one event per 1.7 days of continuous watching,
  a 74-minute sweep would be expected to catch about 0.03 events. Finding none is
  the predicted result, not a clean bill of health. Coverage and exposure are two
  different axes, and a wide sweep buys the first at the cost of the second.

=== WHAT THE SOAK MEANS ===
  This is not a board that fails a memory test. It is a board that passes every
  memory test - including a 96%-coverage sweep - and then corrupts a build hours
  later. That is precisely what makes it expensive to own.
