igneum/docs/plans/release-0.3.12.md

Igneum Miner 0.3.12: the fresh-record rule switch (proving v1) and the app cut, prepared to the publish gate, 6 October 2026

Release engineer, from 08:05 UTC, on the coordinator's instruction ("prepare 0.3.12, APP ONLY, up to the publish gate and STOP there; the project lead gives the go"), widened at 08:55Z on its clock: "no longer app only", the proving agent proved the segment record rule needs a consensus switch, so the node fork proving-v1 0f0dda95 (proving_v1_fresh_rule_daa: never by default, in the digest only once set; from it a fresh segment record is valid whenever the previous segment is not proven) and the app's segment-aligned prover (272b025, docs aea2f6a) ride in it. Worktree /Users/joshm/Projects/igneum-wt-ship0312, branch release-0.3.12 from master ddfcdac; vendor/ symlinked to the main checkout's (46 entries); the fork worktree vendor/igneum-node-0312, branch release-0.3.12-node = 0f0dda95 cherry-picked onto 89dfcb95 (its parent ece42979 is inside 89dfcb95, so the rebase is the one commit: params.rs, exec/proving.rs, exec/rpc.rs, daemon.rs) = 83089544. The 0.3.11 recipe (release-0.3.11.md) throughout; every Mac build under the main checkout's lock; igneum-labs commits. Times are UTC.

1. What 0.3.12 carries

Change	Where	State
/api/state never answers {} again: paid_wei (u128) is a decimal string, the error reply is logged; test	proving-v1 app 6714a45 (the first item)	merged 9bcf4cd (the docs/bench-log.md conflict: both sides kept, the log is append-only)
An update published over an hour before the engine started skips the hourly rollout slot; manifest::unix_from_rfc3339 + tests	update-catchup 2207cd7	merged b984c17
The GPU list ordered by performance (usable, discrete before integrated, rate in 5 MH/s buckets, memory); 3 UI tests	card-order ffb2bfa	merged c6608c1
HiveOS local mode carries the override (OVERRIDE= in the Flight Sheet's extra config, written to data/override-params.json by h-run.sh, C41), rigs mine only until a Linux prover ships, IDENTITIES=auto by VRAM, the per-card README table	hive-words 98271ff (packaging/hive only)	merged b0a6231
Ember Tune: two-knob plans + priors + the UI line; every quit names its source (b671c8b); a second engine never runs the updater (e600e63, C35); no pipe into a second engine (8ab9068); jobrun.rs elevated follow_file (1e9550e); the BOM fix + CI check (8273494); Power control switch (49bbe14 = 3562f26); igneum-gpu-telemetry.exe (ADLX) built by build-windows.sh and carried in the Windows inputs	ember-tune 9a6469f	NOT MERGED: conflicts in seven files against the 0.3.10/0.3.11 app (its base ca8d9f3 predates both): ci.yml, config.rs, engine.rs (the detect path, the power-cap plan, the test module), ui/app.js (four hunks against miner-ui-2's View), ui/index.html (the settings panel 0.3.10 removed), proto-opencl/README.md, bench-log.md. Its agent is rebasing it onto release-0.3.12 (section 2)
The hidden-console builder for every elevated launch, windows-spawn-check.mjs, the PC 1 console-watch scripts	job-console 13755b9 (+ 3562f26 Power control)	NOT MERGED: conflicts in six files (ci.yml, config.rs, engine.rs, jobrun.rs, app.js, index.html), base a93199a; carried by the ember-tune rebase (it already holds 3562f26)
The miner's gRPC resubscribe after a node restart (C43, the 0.3.11 finding)	no commit exists (the ledger entries b19fe5f, 0751dde, c8c831c only)	OWED, listed in section 10
The fresh-record rule switch: proving_v1_fresh_rule_daa (Option, never by default; a node with the field set prints it and carries it in the digest; a fresh segment record is valid from it whenever the previous segment is not proven)	fork proving-v1 0f0dda95 on ece42979	cherry-picked onto 89dfcb95 as 83089544 (release-0.3.12-node)
The segment-aligned prover: a segment record the chain rule refuses is held and offered again every pass until the segment closes; the fast-time harness on the fresh-record rule (both cases); the prover host and export; the WSL2 prover package script; infra/fast-time/override-60x.json (measured by its agent: 9 segments per 30 min on PC 2, 72 of 72 shards paid, 11% hash cost, 17.6 GB peak)	proving-v1 app 272b025 + docs aea2f6a (on 6714a45)	merged 49e0e2c (clean)
The packaged line (C34): the ten-field object of section 4	7dd3ff7	packaged-config.sh --test passes
The six version files	81e4ecb (--check: 0.3.12 in all 6)

Left out on the coordinator's word: prover-floor's server (its packaging row is 0.3.13), explorer d7e797c, pool-v0, rig-install, ota-k2, the ledger forks.

Changelog line (draft, for the manifest notes at the go): "Igneum Miner 0.3.12: the fresh-record rule for proving v1 from DAA 192,000 (a fresh segment record is valid whenever the previous segment is not proven) and the segment-aligned prover; the GPU list in performance order; an old update no longer waits for the hour; Ember Tune (every card tuned for MH per watt, Power control off by default, the app never asks for administrator rights on its own); a second engine never installs over the app; /api/state always answers; HiveOS rigs carry the override. Node 83089544."

2. The branch

Commit	What
9bcf4cd, b984c17, c6608c1, b0a6231	the four merges above, in the coordinator's order (proving-v1 first)
81e4ecb	Igneum Miner 0.3.12: the six version files
ebea8b6	the ember-tune rebase tip 7f6c4e6 (with job-console 13755b9 inside), merged as one branch (section 3)
11e8ca6, ab01f48, 01abcc2	the plan
062c3f8	node-source.pin 83089544 with the second inputs push (the Windows-build commit of 0.3.12)
37b6a7f	tools/proving-v1/pc2-agg-cost.ps1: pkill -f sp1-gpu-server (the CI root-socket check)
88df58e	master d3b64cb merged (docs only): the release tip, CI green
6532adf	make-payload.sh: on CI the AMD telemetry helper is taken from the unpacked inputs (the worker glob igneum-worker-*.exe missed igneum-gpu-telemetry.exe, so the first payload, run 37435975425, shipped without it: the inputs had it, the zip did not). The CI commit of 0.3.12

Checks on 81e4ecb before the rebase landed: the app cargo test --release -p igneum-app under the lock: ok 115 (lib) + 28 (ota-sign) + 8 (prove-verify), 0 failed (08:19:22 to 08:19:28Z, warm target cloned from the 0.3.11 worktree); the UI tests notices, update-card, view: 23 of 23.

3. Builds and artefacts

What	Command	Result
The HiveOS package (first build, app-only cut)	the 0.3.11 node and workers with hive-words' scripts	08:19:45Z: b0a20917... (24,501,272); superseded below once the node changed
The merged tree	ember-tune 7f6c4e6 (release-0.3.12 b0a6231 merged INTO ember-tune as c5918c7, job-console 13755b9 cherry-picked on top; 0.3.11's six-section View and card order kept whole, Ember Tune's TuneLine block and the Power control switch added in 0.3.11's markup, engine.rs keeps the detect arm with the tune fields in hotplug::apply_pref, both test modules, jobrun.rs the hidden-console builder plus follow_file) merged as one branch	ebea8b6, 08:22Z (the CI commit is 062c3f8); the version files still 0.3.12 in all 6; packaged line, node-source.pin and vendor/ untouched against master
The app	cargo test --release -p igneum-app under the lock, then cargo build --release	08:22:56 to 08:23:06Z: ok 133 + 28 + 8, 0 failed; igneum-app 0.3.12 (2,273,664)
The UI and relay tests	node --test notices, update-card, view, tune-line; relay/test/*.test.mjs	26 of 26; 23 of 23
The CI checks on the Mac	identity, no-conflict-markers, copied-sources, signer-pipe, prover-socket, second-engine, bash-body (self-test + tree), kit-path (self-test + tree), windows-spawn (self-test + tree), pinned-guests, no-secrets, check-workflow-shell	all ok; link-check passes after node site/build.mjs (as ci.yml runs it: the committed litepaper.html points at /bench#counter-asic-2-0-the-numbers, an id the site build creates from bench-log.md)
The Windows workers and the AMD telemetry helper	proto-cuda/nvrtc/build-windows.sh (mingw) under the lock, 08:23Z	the worker SOURCES are unchanged against master (git diff master HEAD -- proto-cuda proto-opencl proto-metal igneum-pow: only build-windows.sh, the new gpu-telemetry.c and its .rc), so the inputs carry the 0.3.11-verified workers that mined all night, igneum-worker-cuda.exe 2b3b8c92885442179f6bf2907c6f3eb453dc4a19908d90fd05981a09b7c2674c (1,536,512) and igneum-worker-opencl.exe edc4a75da3b93d814caa69fd635010780d63d5b622ec24c3741d433c584f91e3 (478,208), not this morning's rebuild of the same sources (12bfaa27..., e0fd7042...: mingw PE builds are not byte-reproducible); NEW igneum-gpu-telemetry.exe 8d679b52b19af3cbd6bf4fd6f77d337b2fb78af13d02627e7aa7e92507993459 (387,584; ADLX, SetupAPI, PDH; the Igneum resources, version 0.3.0 as the workers carry)
The Windows inputs	IGNEUM_WIN_RELEASE=<fork>/target-integration/x86_64-pc-windows-gnu/release IGNEUM_NODE_SRC=vendor/igneum-node-0312 packaging/windows/push-inputs.sh, 08:24:29Z (a deploy of the downloads folder only; the manifest untouched)	node 89dfcb95 (igneumd.exe be8e83c0..., igneum-miner.exe 1ba1a249..., PC 2's 0.3.11 build), the two workers above, the telemetry helper, the mingw DLLs and nvrtc; signed, verified, live (HTTP 200); node-source.pin unchanged 89dfcb95
The DMG (first build, app-only cut)	the 0.3.11 node	08:24:33Z: ff630e9d... (41,630,620); superseded below once the node changed
The fork's Mac node, 83089544	CARGO_TARGET_DIR=vendor/igneum-node/target-0312 cargo build --release -j 4 -p kaspad -p igneum-miner --features kaspad/igneum-pow from vendor/igneum-node-0312, under the lock; the target dir cloned by APFS from target-0311; target-integration in the fork worktree links to it	08:38:13 to 08:41:31Z: igneumd 746a931fde9b840ca444a03cd757854e2e2ce7ebc4782ddd00ed161644d705f9 (41,386,160), igneum-miner 5381683e5717d91416c5a97456e0d050dd9645b1e27bce7d55808ce621cc1a26 (8,763,936); igneumd/2.1.0-83089544
The seed's Linux node (glibc 2.36 target, zig)	NODE_SRC=<abs fork> TARGET_DIR=vendor/igneum-node/target-0312-linux OUT_DIR=<scratch>/cross infra/cross/build-linux.sh under the lock	08:38:21 to 08:41:18Z (175 s): igneumd 4f142d5148f218f1286e24e7c4a167aa2f54262336f96e7cf281f520c714fc6f (47,919,144), igneum-miner 38397ae66c265b63db8e5458b46e7feb942121a7dc5625919df0a8d35e7a1ba1 (9,861,072); version.txt names 83089544
The prover host and export (the pinned guests unchanged)	cargo build --release -j 4 -p igneum-prove-export -p igneum-prove-host in proving/igneum-prove under the lock (the target cloned from the 0.3.11 worktree)	08:39:07Z: igneum-prove-host b90d58d0529ce29f0e7ca8ae780a6442f92edcf1c71752fc60fbc72bc5c11fd8 (58,626,560), igneum-prove-export b60056127d32bda363c0e305e73e9f699a5774c0988aee5bfd28a0fc61a56b9a (2,808,160); --mode id: shard program id 0x2b1a81cb413236cf063077b46ed3111628f6c41036bcf6e23ee4cbbf5679ef7a, the pin of 0.3.9 to 0.3.11; pinned-guests-check ok, proving/igneum-prove/elf/ untouched
The HiveOS package	NODE_OUT=<scratch>/cross WORKERS_OUT=<the 0.3.11 Linux workers> VERSION=0.3.12 OUT=<scratch>/hive packaging/hive/make-hive-package.sh (the Linux workers 4aaff27f.../82d90890... unchanged: their sources are)	08:41:54Z: igneum-hive-0.3.12.tar.gz 7972af92e7cd9a032303eca4d95b533f53e0e68d1b9cae5bfe406a5b7c30a454 (24,506,282); h-run.sh writes data/override-params.json from OVERRIDE and starts the node with --override-params-file (the 0.3.11 open item closed); the node inside is 83089544
The DMG	NODE=<fork igneumd> MINER=<fork igneum-miner> PROVE_HOST/PROVE_EXPORT=<this tree's build> packaging/mac/build-dmg.sh under the lock	08:42:49Z: Igneum-Miner-0.3.12.dmg 7a4a5f5f772956e983127280a5ec62a4fcfaf903b3afa38fe3a89a37cee23520 (41,702,535), engine 0.3.12, node 83089544 (igneumd 41,163,744 inside, stripped by the DMG build), the new prover host and export, igneum-bench from proto-metal/main.swift (unchanged, 66ec0e78...), packaged-config carries the ten-field object, hdiutil checksum valid
The node suites with the igneum-pow feature (the coordinator's ask; the PC runner's test units carry no features field, so this is the Mac's run; the PC 2 run is owed to the Counter ASIC 3.0 coordinator's window, section 3a)	CARGO_TARGET_DIR=vendor/igneum-node/target-0312 cargo test --release -j 4 -p kaspa-consensus -p kaspa-consensus-core -p igneum-exec -p kaspa-pow -p igneum-miner -p kaspa-p2p-flows --features kaspa-consensus/igneum-pow,kaspa-pow/igneum-pow from the fork, under the lock, 08:39:31 to 08:45:01Z	igneum-exec 17 of 17, igneum-miner 18 of 18, kaspa-consensus 97 passed, 2 failed, 4 ignored. The two: (1) pruning_proof::igneum_m20_tests::witnesses_are_checked_in_epoch_order_under_their_own_seeds (igneum_m20_tests.rs:122: the expected EpochSeeds.era is all zeros, the code draws 515e...: the test predates the era draw of class v3) FAILS THE SAME on 89dfcb95 (run 08:45:48Z on the 0.3.11 fork): the known M20 era fail, NOT fixed by 0f0dda95, still owed; (2) finality::tests::ban_is_decided_by_the_carrying_block_so_nodes_agree_on_every_voter_list (finality.rs:1883) failed inside the full crate run and PASSES alone on both 83089544 and 89dfcb95: order-dependent, not a regression of this cut, owed as flaky. kaspa-consensus-core 107 passed, 1 failed (config::params::tests::fast_time_60x_file_is_the_devnet_at_60x: infra/fast-time/override-60x.json does not parse into OverrideParams, "duplicate field proving_v1_activation_daa" at line 64: the file has carried a second proving-v1 block since c2544be on 5 October, so the test fails on master's file and on 89dfcb95 alike; not a 0.3.12 regression, the file is owed a dedupe), db_compat 7 of 7, kaspa-pow 14 of 14, kaspa-p2p-flows 33 of 33 (08:47:13Z, no fail-fast). Net: 3 failures, each present on 89dfcb95, none from 0f0dda95
PC 2 build-and-suite job	IGNEUM_WIN_RELEASE=<scratch>/pc2-out node tools/build-job.mjs run --node vendor/igneum-node-0312 --target 1ccfe586 --targets linux,windows --node-tests "kaspa-consensus kaspa-consensus-core igneum-exec kaspa-pow igneum-miner kaspa-p2p-flows" --app-tests igneum-app from this worktree, published 08:54:56Z on the prover-floor agent's "PC 2 is yours" (its floor-core-alone and floor-core-miner closed 08:48:37Z and 08:53:13Z, after the Counter ASIC 3.0 coordinator's release at 08:43:24Z); CPU only, the prover on, nothing else touched. The coordinator's later order (after the prover-floor agent's SECOND pair) arrived once the job had run; the app's queue serialised them anyway: this job ended 09:02:04Z and floor-build-6 started 09:02:05Z, then floor-core2-miner closed 09:13:22Z with the prover on and the miners never stopped, so nothing ran beside a GPU row	build-20261006-085456: started 08:55:24Z, done 09:02:04Z (400 s), every stage ok: Linux node 135 s (igneumd 34877b86..., igneum-miner c779777f...), Windows node 165 s (igneumd.exe 5bbcbd59f592fa31bf31c18516cef81cc0e7e537d398382fc8063e3402d80917, igneum-miner.exe af973318...; PC-built, NOT shipped: the inputs carry the Mac cross-build f580b4aa..., placed under the scratchpad), the app both targets; RESULT test node [the six crates] exit 0 44 s (without the igneum-pow feature, the runner's shape: the M20 era test and the fast-time file test are outside its reach there) and RESULT test app/igneum-app exit 0 6 s
The Windows node exes	CARGO_TARGET_DIR=vendor/igneum-node/target-0312-win proto-cuda/windows-node/cross-build.sh <fork> 4 on the Mac (mingw, the 0.3.5/0.3.6/0.3.9 path; PC 2 is the Counter ASIC 3.0 coordinator's this morning), the target cloned from target-release-win, under the lock	08:40:43 to 08:48:05Z (6 min 42 s): igneumd.exe f580b4aad1e19a47742d0d836a56dad36b9380d3890ca115b1babced4d83a8db (52,177,920), igneum-miner.exe 06c17d4c23c8b1327793bebcd9b2cba115045ea91b68baea8cb92b232a94678b (11,040,768); static (KERNEL32, advapi32, api-ms-win-core only)
The Windows inputs, second push	IGNEUM_WIN_RELEASE=<target-0312-win>/x86_64-pc-windows-gnu/release IGNEUM_NODE_SRC=vendor/igneum-node-0312 packaging/windows/push-inputs.sh, 08:48:28Z	node 83089544 (the two exes above), the 0.3.11-verified workers 2b3b8c92.../edc4a75d..., the telemetry helper 8d679b52..., the mingw DLLs and nvrtc; payload-inputs.zip f8e567bd164b382d32a33fb488df658fa68555092ac6bdac85387d0a8bd5d547 (65,259,161), signed and verified, live (HTTP 200); node-source.pin 83089544 committed as 062c3f8, the CI commit of 0.3.12

| The Windows installer and zip, first runs | runs 37435975425 (ebea8b6, no telemetry helper) and 37436904041 (6532adf, the 0.3.11 node) | superseded | | The Windows installer and zip | windows.yml run 37438673235 on 062c3f8 (dispatched 08:49:12Z after the second inputs push) | Igneum-Miner-Setup-0.3.12.exe f11a296acf1ea3efa8a6151efa357cc5c222e3b2ffec5701ea4bcefe29307810 (45,270,093); igneum-windows-app.zip a6f33ef21bb1d1682f48ca22332d50c0302f4b4f552a99157581f3249c42ea3b (65,507,597): igneumd.exe f580b4aa... (the cross-build, 52,177,920), igneum-miner.exe, igneum-app.exe 0.3.12 (3,700,736), the two workers, igneum-gpu-telemetry.exe (387,584) this time, the mingw DLLs, nvrtc64_120_0.dll and nvrtc-builtins64_128.dll |

4. The override objects and the digests (the 0.3.12 Mac node 746a931f..., ports 60975/60976, 22 s each, under run)

Override file	Lines	Digest
none	igneumd/2.1.0-83089544, no activation line	c562d70e1428c9789823cc40067623b4767f7c555ce7ff4ea11c1498f013ef6c, EQUAL to 0.3.11's no-file digest: the new field is never by default and leaves the digest alone until set
the fleet's live nine-field object	the six activation lines of 0.3.11, no fresh-rule line	0139ab9dc2992d449ec787d8f021974933631eb55740ab4b6ce9d5c226e72888, EQUAL to the fleet's digest today: publish 1 (the binary) changes no handshake, a 0.3.12 node and a 0.3.11 node on the nine-field file accept each other
the ten-field object (proving_v1_fresh_rule_daa 192000)	the six lines plus Proving v1 fresh-record rule from the override file: from DAA score 192000 a fresh segment record is valid whenever the previous segment is not proven	bd786a4b521e87c05bce3da4c46b4f4696deb16dfbdc913f181c980a8eb51688

The ten-field object (the packaged line 7dd3ff7, the manifest of publish 2, the hand nodes' and the seed's files at step 2):

{"difficulty_v2_activation_daa":33000,"proving_v0_activation_daa":84100,"fees_v1_activation_daa":210000,"finality_v3_activation_daa":135200,"program_class_v3_activation_daa":154800,"proving_v1_activation_daa":154800,"proving_v1_segment_blocks":8,"proving_v1_unproven_daa":600,"proving_v1_aggregator_share_bps":1000,"proving_v1_fresh_rule_daa":192000}

N = 192,000 is tip + 14,400 for a publish near 10:15Z (tip 171,933 at 08:38:41Z, 0.97 to 0.99 DAA/s, so about 177,600 then); the floor N - tip >= 10,800 holds while the tip is at or under 181,200 (about 11:15Z); past it N is re-pinned to tip + 14,400, the packaged line, the DMG and the installer rebuilt (CI again), and the digest re-read. A 0.3.11 node given the ten-field file dies on the unknown field (deny_unknown_fields), which is why publish 2 comes only after every node runs the 0.3.12 binary (the reviewer's C39, the 0.3.11 order).

5. The publish gate: what runs at the project lead's go, in which order (the 0.3.11 two-publish shape)

Nothing below has run. Runbook: the session scratchpad's r0312/rollout-0312.sh (every step a function; step_floor before each publish).

Step	What	Gate
0 the floor	step_floor: 192,000 minus the tip's DAA at least 10,800	read before publish 1 and again before publish 2
1a the hand nodes, the seed	the observer and node 1 on the 0.3.12 binary with the NINE-field file (IGNEUMD=<fork>/target-integration/release/igneumd IGNEUMD_COMMIT=83089544 infra/devnet/restart-hand-nodes.sh '<nine>'), then the seed (IGNEUMD_LINUX=<scratch>/cross/igneumd IGNEUMD_LINUX_SHA256=4f142d51... infra/devnet/restart-seed.sh '<nine>'); every one prints 0139ab9d..., nobody is refused; then step_mac_miners (the Mac's miner does not reconnect to a restarted node 1 by itself, C43)	the project lead's go
1b publish 1	node tools/ship-app.mjs 0.3.12 --node vendor/igneum-node-0312 --branch release-0.3.12 --public --activation-height 154800 --deadline-note "program class v3 + proving v1" --notes '<section 1>' --from ci: ci "already" (the green Windows run), fetch, dmg "already", copy, manifest with consensus CARRIED OVER (the nine-field object; the digest stays 0139ab9d...), deploy, verify (--from console after the public index settles at the edge), the console item; --public carries the HiveOS package 7972af92...	after 1a
1c update-now	the Mac (d937c69d) first; the laptop (37ba0461) with it if it is on the air; PC 2 (1ccfe586) on the Counter ASIC 3.0 coordinator's word (PC 2 is its this morning; the proving agent's constraints: the app's prover stays on, no quit or restart of anything but the update's own); PC 1 (ae432dc7) last, once the project lead has relaunched its app (down since 22:31:06Z yesterday, on 0.3.10: it takes the nine-field object and 0.3.12 at its relaunch through the manifest; Power control is off by default so nothing asks for administrator rights)	each machine's STATUS line back on 0.3.12 with 0139ab9d...
2a the floor again	step_floor	>= 10,800 or re-pin
2b the hand nodes, the seed	the same two scripts with the TEN-field file; each prints bd786a4b... and refuses the nine-field side until it switches; step_mac_miners again	every app node on the 0.3.12 binary (1c)
2b publish 2	publish-manifest.sh --version 0.3.12 --override '<ten>' --activation-height 192000 --deadline-note "proving v1 fresh-record rule" --notes '<section 1>' --public --deploy	after the hand nodes
2b update-now (switch)	the Mac and the laptop, then PC 2 on the 3.0 coordinator's word, then PC 1: each app writes the ten-field file at the manifest take and restarts its node at a safe moment (the Mac's node is node 1, already switched: nothing to restart)
the sweep	every node prints bd786a4b521e87c05bce3da4c46b4f4696deb16dfbdc913f181c980a8eb51688; the fresh-record rule arms at DAA 192,000

One line for the project lead, per machine, when he says go: the Mac and PC 2 each restart their engine once for 0.3.12 (under a minute, the miner back on the next template) and their node once more for the fresh-record switch (a few seconds, mining resumes on the same chain); PC 1 does the same at its relaunch and, with Power control off by default, never asks for administrator rights again (its 5090 runs uncapped until he switches Power control on in Settings); the observer, node 1 and the seed are restarted by hand twice; from DAA 192,000 (about 14:20Z) a prover may file a fresh segment record whenever the previous segment is not proven, so paid segments stop stalling behind an unproven one; until the switch nothing changes in consensus (digest 0139ab9d... through publish 1).

6. CI

Run	On	Result
ci 37435705568	ebea8b6 (the branch push, 08:22:40Z)	green (pow tests and census build, simulators, site build + link check + identity, the PowerShell 5.1 parse job)
windows-ci 37435975425	ebea8b6	green 08:30:58Z (the parse job 08:25:16 to 08:25:56Z; engine, window host, payload, installer, smoke run 08:26:00 to 08:30:58Z); superseded by the run below (no telemetry helper in its payload)
ci 37436904569	6532adf (the branch push, 08:33Z)	(pending)
windows-ci 37436904041	6532adf	green 08:39:20Z; superseded by the run below (the node changed)
ci 37436904569	6532adf	green
windows-ci 37438673235	062c3f8 (gh workflow run windows.yml --ref release-0.3.12, 08:49:12Z, after the second inputs push)	green 08:54:27Z (the parse job 08:49:18 to 08:49:59Z; engine, window host, payload, installer, smoke run 08:50:02 to 08:54:27Z against the 83089544 inputs); fetched 09:03:17Z with OTA_SKIP=1 CONSOLE_SKIP=1 packaging/windows/fetch-ci-artifacts.sh 37438673235 into the downloads folder, NOT deployed
ci 37438674529	062c3f8	FAILED in one step, prover-socket-check.sh: tools/proving-v1/pc2-agg-cost.ps1 (in through the proving-v1 merge) ends its root prover with pkill -x sp1-gpu-server, and the check wants pkill -f; the line now reads pkill -f ... ; rm -f /tmp/sp1-cuda-*.sock (one playbook line, nothing the app or the packaging reads; git diff 062c3f8 <fix> -- app packaging proto-cuda proto-opencl proto-metal vendor is empty, so the Windows artefacts of 37438673235 stand, as 0.3.11's did across 3b0262f and 2a62735); the rerun is the row below
ci 37440456687	37b6a7f (the one-line playbook fix)	green 09:07Z
ci 37440559793	88df58e (master d3b64cb merged in: the morning summary, docs only; the release tip)	green 09:08:08Z. The 0.3.12 CI verdict is therefore run 37440559793 on 88df58e; the Windows build is run 37438673235 on 062c3f8, the same app, packaging and node sources (git diff 062c3f8 88df58e -- app packaging proto-cuda proto-opencl proto-metal vendor igneum-pow is empty)

The ship state file ~/.cache/igneum/ship/0.3.12.json carries sha = 062c3f8 (the Windows-build commit, which the ci step looks up by commit; the tree is 88df58e, docs and one playbook line later, as 0.3.11's was two docs commits past its build commit), forkCommit 89dfcb95, bumpedAt before the DMG's mtime (so the dmg step reads "already"), and runId once the Windows run is green.

7. Owed to the next cut (0.3.13)

Item	What
C43, the miner's dead gRPC channel	no commit exists; igneum-miner mine grpc:// must re-subscribe after its node restarts (the 0.3.11 finding: 11 min of 26 MH/s burned on the Mac); until then every hand restart of node 1 is followed by restart --what miners
prover-floor's server	its packaging row is 0.3.13 (the coordinator's word)
explorer d7e797c, pool-v0, rig-install, ota-k2, the ledger forks	left out on the coordinator's word
the Windows workers' reproducibility	mingw PE builds differ byte-for-byte between builds of the same sources (12bfaa27 vs 2b3b8c92 today); a -Wl,--no-insert-timestamp (or SOURCE_DATE_EPOCH) in build-windows.sh would make G5's one-commit rule checkable by hash
the site build's non-idempotence and the pre-push flip	unchanged from 0.3.10 and 0.3.11