=== allowed, executed, succeeded, still wrong (exit 5) ===============
    scale_pool asks for 60. The envelope allows up to 64 - this call
    is inside the envelope's own bound. The database this pool talks
    to has max_connections=56, so the target CLAMPS and returns
    success anyway. STEP 6 asks whether pool_max equals the value
    that was REQUESTED, not whether the call succeeded, and it is the
    only step that catches what the target itself did not report:
    allowed, executed, succeeded, still wrong.

guarded run RUN-FD-CLAMP-01  agent=meridian-ops-agent  target=fielddesk-api/control  transport=local
  clock 2026-08-18T15:25:00Z  policy envelope/meridian-ops@2026-08-29.2  request REQ-39245E8D6B
  STEP 0 prereq       no handoff declared for this envelope: meridian-ops-agent is an ops automation account, not a registered AI system on Meridian's register, so it has no AI Assurance handoff to read.
  STEP 1 envelope     ALLOW scale_pool  class=reversible  zone=it-data  (max_conns in [8,64])
        budget: 0/6 actions, 0/3 mutating used in RUN-FD-CLAMP-01
  STEP 2 pre-state    captured dbf3098bde64c1db  pool_max=32 version=1.14.3 backups=3
  STEP 3 dry-run      pool_max 32 -> 56 (requested 60, clamped at the database ceiling)
        plan hash 74e76c3a5df3b13b - this is what an approver signs
  STEP 4 approval     not required (envelope decision: allow)
  STEP 5 execute      pool_max 32 -> 56 (requested 60, clamped at the database ceiling)
  STEP 6 verify       FAIL  pool_max is 56, the action asked for 60
  STEP 7 compensate   scale_pool {"max_conns":32}: pool_max 56 -> 32
        pre-state restored: yes (dbf3098bde64c1db)
  STEP 8 receipt      seq 1  44826b2d73a50928  prev GENESIS
  STEP 8 receipt      seq 2  bfad32e2900d6a50  prev 44826b2d73a50928  (compensation)
RESULT compensated  exit 5
