release-0.3.20 plan: section 33, b7cc37e7 struck (the 10 s stop), the pin c4459193, the carry ruling, main's F8 and slot rulings, the 12 GB rule in and green, the roll complete, the server-architecture fact, the 0.3.21 driver-check tip

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

docs/plans/release-0.3.20.md

@@ -401,3 +401,17 @@ Speculative builds on the candidate from the ship worktree, box only, sequential
 **Main (13:2x BST):** the driver check (the Intel lane's b9487dcf on driver-check: the table, the install thread with one elevated prompt, the manifest's drivers object, the row strip; box 219 + 32 + 8, cross green, pre-push 35, UI 32) rides 0.3.21 with the drivers manifest once the NVIDIA and AMD hashes are real; 0.3.20's app tree is closed. Clock: the node lane's "UK" stamps read an hour fast (UTC+2); every estimate is re-read against `TZ=Europe/London date` before it is quoted; the node lane and the fleet told to quote that clock. The 15:30 BST checkpoint is the real clock.
 
 **The fleet's prover roll, p1-3080 (13:1x BST):** the pair right (28 "pair ok"), 0 paid in its history (2,167 claims), every proof since the pair dies at the compressed step at the memory wall (device_used 9,859 of 9,885 MiB), 29 of 29, miner off the proving step. Per tier: a 10 GB card does not prove on this pair; the failing prover took the 3080's GPU from its miner for 16 hours for nothing. On the shipper's word (fleet config, no key moves): PROVER=0 on p1-3080 with its lock line and the miner's rate before and after; the same reading for standing boxes under 12 GB; one rented 3080 for an hour to measure the lower-memory SP1 threshold. For main: the kit and the app refusing to start proving under 12 GB with the reason shown, until the measured threshold lands. hub-1 is the roll's last box.
+
+## 33. The pin moves to c4459193; the 12 GB rule in the app; the roll complete (13:3x BST, `date`)
+
+**b7cc37e7 struck:** its digest gate passed (12:14 to 12:16Z: thirteen fields a89be8a7 on both binaries, the sixteen-field object db9a85f9 refused, the live file's own digest eada4bda as 5899f603 reads it) and the fleet's kept-datadir read passed on p12-vast (12:17Z: "Virtual state: the row was written by a build before the silence field (a kept datadir); read as the v1 layout and rewritten in the current one (1 mergeset rewards)", the finality blob converted, 1,747 locks, no panic; a second start first-try; 6a3432a3 the known-failed shape), but its ten-minute mixed-version gate FAILED at the restart step (12:24:19Z): the node died on its own datadir LOCK (conn_builder.rs:167) because the 6a3432a3 watchdog sleeps its whole 10 s poll before checking shutdown, so every node from 6a3432a3 on takes up to 10 s longer to stop than 5899f603 (the fleet's "a 12-second timeout does not stop the node" was this) and a restart inside that window meets the lock. Fix c4459193 (b7cc37e7's child, rpc.rs alone): the poll in 250 ms steps returning the moment shutdown is set; test `a_shutdown_returns_within_a_second_whatever_the_poll`; exec suite 31 passed with the three watchdog tests. **The candidate pin is c4459193**, igneum-pow 8c728ca3; its build on build-1 from 12:30Z, digest and ten-minute gates on its own binary, lines about 12:50Z (13:50 BST). Carry ruling (shipper): the b7cc37e7 kept read is evidence, not the gate line; the gate line is the kept read on c18-1 on c4459193's binary before the canary's restart step. The b7cc37e7 pairs (hands bc6b3b3b / 240eb0a4, seed ffaa441d / 64207cc0) are evidence only; the build-server lane rebuilds on c4459193 and runs the PC 2 Windows kept-datadir job on c4459193's exe.
+
+**Main (13:2x BST):** the attack-pass lane's F8 re-gate on object 5 heads to FAIL at 1.2x on nine of the first thirty seeds (worst p31 at 29.3x), far better than the old stream; 0.3.20 ships object 5 as it stands (the live floor flips every node to the OLD stream on 13 October otherwise); sub-version 2 on the Counter lane is 0.3.21's; the 16:00 report tells the project lead the floor move is recommended, not optional. Build slots: suites to build-2, builds and gates on build-1 (tools worktree fast-forwarded to master 1cf850c9; build-remote routes test and bench to box 2 by class).
+
+**The 12 GB rule in the app (4c89372a):** MIN_VRAM_MB_PROVE_ANY 11,800 and PROVE_UNDER_12GB_LINE "proving needs a 12 GB card; mining continues" in provedefault.rs with `the_prover_refuses_every_nvidia_card_under_12gb_and_says_why`; the prover loop refuses before the sync wait when every present NVIDIA card is under it (status off, the sentence); the tile sentence in app.js (PROVE_MIN_GB 12) with its view test. App gate on the box: 215 + 28 + 8, rc=0 (13:13 BST); UI 31 of 31; the Windows cross of the app running. In the fleet's kit: box-prover refuses the same way (gpu-fleet f7d40af5, on 18 boxes at 12:14Z; p1-3080's line "RESULT refused 2026-10-07T12:14:30Z proving needs a 12 GB card; mining continues (memory.total 10240 MiB)"). PROVER=0 on p1-3080 (12:13:57Z, node and miner untouched, no key moved; lock line 98.6 percent at checkpoint 8931): the miner 38.5 to 39.8 MH/s before, 44.7 to 45.9 after, plus 17 percent.
+
+**The prover roll complete (12:19:40Z):** 11 of 13 standing provers paid on the 71bc2438/263bf4ce pair (hub-1 last, 4.6410 IGN after 143 s); p2-4070-1 pair ok and unpaid (the claim race), p1-3080 prover off. Feed 12:25Z: provers_10m 5, 67 shards, lag 351 s.
+
+**Prover kit facts from the 3080 hour (the fleet, corrected 12:3xZ):** every standing box's sp1-gpu-server is a binary built for its own card (p1-3080 sm_86, p2-3090-3 sm_86, p1-4070 sm_89, p1-5090 sm_120); a server for the wrong architecture fails every proof in 12 s with "CudaRustError: named symbol not found" and the miner never notices. Kit consequence (0.3.21 design, routed to main): one server per architecture picked by compute capability at install, or a fat binary. The 610-driver question is open again (the first host had both faults at once); one more rented 610 host with an sm_86 server separates them. The threshold series restarted on thr-3080b with the sm_86 server.
+
+**For 0.3.21 (the Intel lane):** driver-check tip 1982dbc7 on release-0.3.20's 4c89372a (the stray #[test] at detect.rs:990 removed; the table measured: NVIDIA 617.42 f115c927…, AMD 26.9.2 593c1d73… with the https://www.amd.com/ referer, Intel 32.0.101.9034; box 219 + 32 + 8, gate 35, UI 32).