d-miner5(kat): pin throttle re-fire after window elapses (#567)

* d-miner5(fix): gate back_online on a reported outage

evaluate() emitted a back_online event whenever the prior sample was
offline, regardless of whether an OFFLINE alert had actually fired. A
sub-threshold blip (dark < offline_min) never fires OFFLINE, so the
recovery edge would tell the miner they are back from an outage they
were never told about and that never met the alert threshold -- a
false-positive notification.

Gate the back_online edge on offline_fired so a recovery is only
reported when the outage itself was reported. Add KAT scenario 6b
(blip-then-recover stays silent); selftest 10/10.

* d-miner5(kat): pin hashrate_drop honesty (no fabricated drop without observed peak)

Extends the D-MINER.5 notify-engine selftest to regression-lock the
HASHRATE_DROP edge cases, completing the per-kind honesty sweep
(offline/back_online/daily-summary already pinned):

  6c first-online (peak==0 guard) must NEVER fire a drop -- a drop with
     no observed prior peak is a fabricated alert.
  6d a real halving vs the observed peak fires exactly one drop.
  6e a dark (hashrate==0) sample is an OFFLINE concern, never a -100%
     drop -- drop is evaluated in the online branch only.

Test-only / SAFE-ADDITIVE: pins existing honest behavior, no engine
logic change. Selftest 13/13.

* d-miner5(kat): pin undelivered-no-throttle (a failed alert must retry, never masquerade as handled)

* d-miner5(kat): pin throttle re-fire after window elapses (stale alert must never masquerade as permanent coverage)

Completes the throttle-honesty triad with #8 (within-window suppress) and
#8b/#563 (undelivered retries): once the throttle window has ELAPSED, a
still-true condition (e.g. an ongoing outage) must re-fire. Guards the
`< window` bound in recently_sent() against a >=/off-by-one regression that
would swallow every later real alert and warn the miner only once though
the rig is still down. selftest 15/15.

---------

Co-authored-by: frstrtr <frstrtr@users.noreply.github.com>

Author: frstrtr

scripts/miner_notify_engine.py

@@ -455,6 +455,22 @@ def put(ts, worker, hr):
     check(rt and rt["status"] == "deliver",
           f"undelivered must not throttle retry, got {rt}")
 
+    # 8c) throttle honesty (the partner of #8/#8b): once the window has
+    #     ELAPSED, a still-true condition must re-fire. A throttle that never
+    #     releases would let a single stale alert masquerade as permanent
+    #     coverage of an ongoing outage -- the miner is warned once and never
+    #     again though the rig is still down. Guards the `< window` bound in
+    #     recently_sent() against a >=/off-by-one regression that swallows
+    #     every later real alert.
+    con.execute("INSERT INTO notifications(ts,worker,kind,severity,route,status)"
+                " VALUES(?,?,?,?,?,?)",
+                (t0, "w6", OFFLINE, "warn", "x", "delivered"))
+    con.commit()
+    refire = route_event(con, {"worker": "w6", "kind": OFFLINE,
+                               "ts": t0 + sub["throttle_s"] + 1, "detail": "d"}, sub)
+    check(refire and refire["status"] == "deliver",
+          f"condition must re-fire after throttle window elapses, got {refire}")
+
     # 9) no subscription -> route_event returns None (logged undelivered, not dropped)
     nosub = dict(sub, enabled=0, channels=[])
     check(route_event(con, info_ev, nosub) is None, "unsub should route to None")
@@ -465,7 +481,7 @@ def put(ts, worker, hr):
         for f in fails:
             print("  -", f)
         return 1
-    print("SELFTEST OK (14/14)")
+    print("SELFTEST OK (15/15)")
     return 0