Pool vardiff: the idle easing no longer consumes the sized first correction; docs/plans/pool.md 9.3: the re-run's close and the member rows

pm-1, 7 October 2026: the member was idle for four intervals while its worker compiled, the first-phase idle easing
set first_done, and the measured 190 shares a second then walked down one step per 30 s for 5.5 minutes (the share
check at 10 to 11 ms on pool-1's cores). The sized correction now stays owed through idle easings (test
an_idle_easing_does_not_consume_the_sized_first_correction). Section 9.3: 207 found / 144 confirmed / 56 own orphans
under 2f6c0358, the residual as the node's 1 to 1.6 s template latency (a node-lane row), the member findings.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs/plans/pool.md

@@ -289,3 +289,40 @@ Consequences per tier: a pool operator needs a box whose memory serves the 1 GiB
 verifies about 100 shares a second per core at 9.8 ms; at one share per 10 s per member that is 1,000 members per
 core, so the verifier is not the limit once vardiff holds); a member on any card is unaffected by any of this; the
 chain walk now costs the node at most 600 `get_block` calls per 5 s on its own connection.
+
+### 9.3 The re-run's close (06:33:33Z) and the member rows from its logs
+
+After 2f6c0358 (06:20:34Z to 06:33:33Z, 13 minutes, four members): 207 blocks found, 144 confirmed, 69 orphaned (13
+inherited pending at the swap, 56 on its own jobs: a 27 percent residual at 1 bps), 20,404 shares accepted, 0 rejected,
+0 mismatches, 0 refused, every STATUS `node ok`, the gRPC connection churn gone (ids #3 to #5 once each in a minute; the
+193 connections in 150 s were under f808b3f3's pruning-point walk), re-subscribe lines 2 (one per daemon start: the
+re-subscribe loop was never the churn). Before it, f808b3f3 had worked through the walk by about 06:14Z: 245 blocks,
+69 confirmed, 163 orphaned. First CONFIRMED line 06:24:32Z, 4.80 IGN to the finder. Logs: `~/Desktop/fleet/pool-run-0607/`.
+
+The residual orphans are template latency: the daemon's STATUS reads `last template 1 s ago` and the solo miner on the
+same node reads `template_ms=1632`; pool-1's node answers a template request in 1 to 1.6 s although its mining manager
+answers a per-member request from its cache by a coinbase rewrite (`modify_block_template`) and its prewarm builds in
+0.6 ms. At 1 bps a template that arrives 1.6 s old loses about a quarter of the blocks found on it, pool or solo. That
+latency is the node's RPC path, a row for the node lane (measure `get_block_template` round trips on a devnet node under
+a miner and a pool; the suspect is the RPC service's `pow_epoch` derivation per request). 82 "RPC request timeout" lines
+after the 2f6c0358 swap are the gRPC client's own 5 s request timeout on that path; `--template-timeout-s` cannot
+lengthen it.
+
+From pm-1's log (1.09 GB): 4,342,515 lines are `worker: error N epoch seed mismatch: this worker holds epoch eea66ce7...`,
+one per re-queued job, for the 40 s the worker compiled the new epoch's pack (NVRTC 40,327 ms) and again after every
+reconnect, because a session respawned the worker (a second process beside the first, the same compile again) and
+the member fed jobs for a pair the worker did not hold yet. Vardiff from the same log: shift 10 to 11 (idle easing
+while the worker compiled, which consumed the sized first correction), then 11 down to 0 one step per 30 s over 5.5
+minutes at about 190 shares a second, the load that put the share check at 10 to 11 ms on pool-1's cores.
+
+Fixed in the member (fork commit after b8070476): one worker process per run, jobs held for a pair whose prepare is
+pending and never re-prepared once held (`WorkerMemory`, test `jobs_are_held_while_a_pair_is_being_prepared`), worker
+error lines summarised (one per class, then a count per 1,000), `set_target` applied to the current work at once.
+Fixed in the pool (`vardiff.rs`): the idle easing no longer consumes the sized first correction (test
+`an_idle_easing_does_not_consume_the_sized_first_correction`).
+
+Consequences per tier: a member on a 3070-class card loses 40 s of hash at every epoch roll to the NVRTC compile unless
+its pack is prepared ahead (the pool's `seeds` line carries `next_epoch_seed` for that, the member's prepare-ahead is
+the next member row); a member on a reconnect now keeps its worker and its compiled pair; a pool operator's verifier
+sees the sized correction within 10 s of a member's first shares (one share per 10 s per member from then on) instead of
+5 minutes of a 190-share-a-second flood; the 10-member load figure is still owed and now has a clean form to run in.

pool/src/vardiff.rs

@@ -79,8 +79,11 @@ impl Vardiff {
                 return None;
             }
             if n == 0 {
-                // nothing yet after an interval: wait up to three intervals, then one easier step
-                if (now_s - self.started_s) < 3.0 * self.interval_s {
+                // nothing yet after an interval: wait up to three intervals, then one easier step. The sized
+                // correction stays owed (first_done stays false): 7 October 2026, pm-1's worker compiled its pack for
+                // 40 s, the idle easing consumed the first correction, and the measured 190 shares a second then
+                // walked down one step per 30 s for 5.5 minutes, saturating the verifier
+                if now_s - self.last_change_s.max(self.started_s) < 3.0 * self.interval_s {
                     return None;
                 }
                 self.shift = (self.shift + 1).min(cap);
@@ -92,8 +95,8 @@ impl Vardiff {
                 } else if per_interval < 0.66 {
                     self.shift = (self.shift + steps).min(cap);
                 }
+                self.first_done = true;
             }
-            self.first_done = true;
         } else {
             if since_change < self.min_change_s {
                 return None;
@@ -226,4 +229,23 @@ mod tests {
         }
         assert_eq!(v.retarget(t, 1 << 20), Some(2), "floor at min_shift");
     }
+
+    /// 7 October 2026, pm-1: a member idle for four intervals (its worker compiling) gets eased one step, and the
+    /// first measured rate must still size the jump (190 shares a second at shift 11 is 11 steps away from one per
+    /// 10 s; the sized step takes 8 at once), not walk down one step per 30 s.
+    #[test]
+    fn an_idle_easing_does_not_consume_the_sized_first_correction() {
+        let t64 = 1u64 << 36;
+        let mut v = Vardiff::new(10, 0, 20, 10.0, 0.0);
+        assert_eq!(v.retarget(31.0, t64), Some(11), "idle three intervals: one easier step");
+        assert_eq!(v.retarget(40.0, t64), None, "still idle, nothing within the next three intervals");
+        // the worker is ready: 190 shares a second for 10 s
+        let mut t = 40.0;
+        for _ in 0..1900 {
+            t += 10.0 / 1900.0;
+            v.on_share(t);
+        }
+        let s = v.retarget(t + 0.1, t64).expect("the first measured rate sizes the jump");
+        assert!(s <= 3, "sized first correction from 11 took at most 8 steps at once, got shift {s}");
+    }
 }