release-0.3.20 plan: the cases rerun on c4459193 (the class byte touch), the 610 read (the driver proves; the server per architecture is the rule)

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

docs/plans/release-0.3.20.md

@@ -429,3 +429,7 @@ Speculative builds on the candidate from the ship worktree, box only, sequential
 **CASES END, PASS on dc141409 (the fleet, 12:37:18Z; run from 11:18:10Z):** relay c17-relay (the 0.3.17 relay build, pool-1's kept 1026 datadir, 22 peers, 1,038 blocks relayed) and poison c17-poison (the poisoned copy, mining from 12:24:44Z) against c18-1 (dc141409): over 13 polls, poison wbv 0 and got 1 (one block at first contact, nothing after, the hub refused it), relay R_wbv 0, c18-1 T_wbv 0 throughout, T_from_relay 0, T_1026 headers seen and none accepted, blocks moving with the tip, powEngine igneum-pow; the hub holds 900 of c18-1's blocks in its last 900 and names it in 1 reject (the right outcome). The isSynced flap visible on this line (4 of 6 polls false at the tip), fixed from 8097d600. The three pods destroyed. The ten-member pool window rented at 12:37:36Z. The carry to c4459193 goes to main on the node lane's diff argument (every file from dc141409 to c4459193 against block and header acceptance, the version check, relay and peer handling); if any touches them the cases rerun on c4459193 beside c19-1 whatever the clock.
 
 **The wipe canary started (c19-1, 12:38:55Z):** the node on a wiped datadir on c4459193 (sha256 45be9b02d1b002f5486d0f0108571c3b6042094113ad9da6f3d3d9ffc0072bba asserted before the put; the node's own line igneumd/2.1.0-c4459193, digest eada4bda8aa8368c, finalityBehind false, pruningDepth 108000, 4 peers at 12:39:26Z, IBD from the pruning-point proof), the exec poller every 5 s, the miner c7cfc40b staged. Clock by the 98-minute class: synced about 14:17Z (15:17 BST), the ten-minute mining read and the hub read to 14:30Z, the restart read to 14:40Z, the kept read on pool-1's copy right after. Polls every four minutes.
+
+**The cases rerun on c4459193 (13:4x BST):** the node lane's diff dc141409 → c4459193 (13 files, +1,155 −48) touches one of the cases' four categories: consensus/core/src/igneum.rs sets CLASS_SIGNAL_V4 4 → 5 (the node stamps byte 5, the tally counts byte 5 and above); header_version_acceptable, signalled_version, class_signal_of and the IBD guard untouched; flow_context.rs keeps the relay broadcast's place and payload (submit_rpc_block returns after the block task, on_new_block spawned); nothing on inbound relay or peer paths; the rest is finality, stores, exec RPC and tests. By rule the cases rerun on c4459193; the expected shape is dc141409's with the header byte read 5 (the void 8097d600 run showed the old hub taking 235 byte-5 headers, 0 rejected). To land inside the checkpoint the fleet runs it now beside c19-1: the target a fresh pod with c4459193 on pool-1's kept 0.3.17 copy (synced in minutes by the N13 path, itself a kept-datadir read on the pinned binary), the relay and poison pods against it, CASES END about 15:05 BST if the pods are up by 12:48Z.
+
+**The 610 read (the fleet, 12:40:14Z):** thr-610 (RTX 3080 10 GB, driver 610.57.04, the sm_86 server, the same segment) completes the chain proof, rc 0, 87 s, device_used 8,729 MiB. The 610-series driver proves; this morning's "named symbol not found" was the architecture alone (an Ada server on Ampere cards). The kit's driver rule keeps its floor (570 or newer, CUDA 12.8) with no upper bound; the rule that matters is one GPU server per compute capability, picked by nvidia-smi compute_cap at install (a mismatch fails every proof in 12 s and the miner never notices): 0.3.21's kit design for main. The 10 GB tier unchanged (proves alone at 8.6 to 8.7 GB, never beside its miner; 12 GB fits both with 2 GB spare).