- spec 3.10: C5/3.8 (min_daa = weight window, window-filling report), Q4 (drawn aggregator at
once, fallback for anyone), S1 (draw by weight), 3.9 (finality_reason) rows for fin-fixes da1eb889
- fork-divergence: four rows for the fin-fixes files and the merge note against the difficulty
branch (hot swap is already in master)
- bench-log: unit tests and the scenario 2 and 5 re-runs before (master) and after (fin-fixes)
- fud-ledger: F17 and F1 status lines
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simulator harness over sim/difficulty/sim.py with multi-miner attribution and in-rule timestamp forging, a 3-node CPU test network (ports 27700+), results and bench-log entry. Timestamp stretching inside Kaspa's rules drops the Igneum block rate 34 to 88% (the per-step clamp cancels forged and honest pairs to zero time); proposed 10 s timestamp bounds plus a sanitised running clock in the chain steps (+0.7% to +1.1% drift at 50% in the simulator). Block flood underflows the 192-bit work after 4,142 blocks; a 2^128 target floor proposed. Rule not changed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
tools/finality-attacks: run.mjs drives a private igneum-devnet-800 network (ports 27800+,
/tmp/igneum-fin-attacks, skip_proof_of_work) with the test-only hostile flags of igneum-miner
(vmine, --equivocate, --sybil, --drop-votes, --pulse, fin-rpc-attack; worktree fin-attacks on
master c6d47547..2a00ff55). Seven scenarios in priority order 3,2,1,6,4,8,5 with a spec 03
criterion and a measured result each; README carries the catalogue, what needs a finality-aware
p2p probe, and a proposed diff for every FAIL.
Results (six voters): S3 dishonest aggregators PASS (35/35/35 locks, 0 conflicts, 1,018 ms);
S2 Sybil dust: weights PASS, aggregator sortition FAIL (per key, ledger F17); S1 equivocation
PASS (2/2/2 stripped, 0 conflicts); S6A 3/3 partition FAIL (floor is time-bounded, one side
crossed 56.7% at 84 s of a 90 s split, T* = 2F/13R, 9.2 days for a 50/50 split at mainnet
scale); S6B 4/2 PASS; S4 vote-dropping producer PASS (0 ms added); S8 malformed votes over
RPC PASS (9 cases, no crash); S5 pulse: no retarget amplification (ratio 0.999) PASS,
lock-alone FAIL (a 20 s burst locked checkpoints 1 to 10 alone on a young window, ledger F1,
spec 3.8 not implemented). S7 eclipse not run.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
No per-job growth in any worker or in the miner's memory. The STATUS rates are cumulative averages
(a fast first interval decays by construction), and the miner's Seeder walks the selected chain from
the sink to the epoch start on every memo miss (one getBlock per block, up to 3,600), a gap between
jobs that grew 0.10 s to 0.33 s across epoch 2 on the PC and reset at the epoch boundary while the
difficulty held. Reproduced on the Metal worker (churn on, off, on). One versus eight workers at two
fixed difficulties: 4% and 1% constant cost, no decay. Fix as a unified diff, not applied.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
sim/economy/sim.py: 1,000 operators choosing MINE, PROVE, HYBRID or OFF per card class with their own clients; sortition by weight with the 10-s window then open claiming, external jobs with the 90/10 split, backlog rule, difficulty clamps, GBM price. Scenarios a to f, 5 seeds: no backlog, no window miss, hash floor 0.74 of pre-event. Traffic sensitivity finds the shortage oscillation only above the proving fleet's capacity (100 to 300 shards per block); at 100 the sortition window (10 s to 20 s) is the lever that removes it. docs/analysis/economy-2026-10-04.md holds the model, assumptions, results, worst case and the proposal (window = p90 shard time plus a swap, 25 s at today's targets; B_p tied to the live fleet), not applied.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Adversarial robustness and conformance tests of the execution layer against a
throwaway 3-node simnet on ports 27600+. Six scenarios, each a runnable command
with a design-derived pass criterion and a measured result: malformed/boundary
txs, nonce games across parallel blocks, RPC fuzz, pgas exhaustion, reorgs under
execution, and developer-registry abuse. 98 checks, 0 failures, 0 node panics.
Two findings filed in the bench-log entry: the mempool admits txs with gas_limit
above B_e (low), and an over-pgas-budget tx is executed natively in full before
being skipped for no fee (medium, griefing).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Safety: X = 1/3 of total weight while honest votes reach every honest node within 41 min of median time; 4/30 = 13.3% across a longer partition, because only the 17/30 floor binds then (2 x 17/30 - 1). Liveness: Y = 17/30 connected and signing, T = P (1 - Y/(2(1 - Y))) + 107 s, 107 s at 2/3, 43 min at 17/30; below the floor finality pauses and the node reports it. Two certificates at one index: no verified lock is ever withdrawn, operators resolve (replaces the 3.5 re-evaluation). Seeds: uncertified checkpoint allowed (O-4.3 decided). Scenarios H (equivocator across a 50/50 split: conflicts at 12 to 16 min with 20%, none at 13%), I (40/40/20), J (signing stops 1, 6, 24 h), K (bought keys worth 20% and 40% against 30% hash), five seeds each. 3.10 rows for the gaps; open items O-3.15 to O-3.19.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/evidence.md and site/evidence.html: 28 public claims with one of five status labels (7 designed, 3 implemented, 18 tested by the team, 0 reproduced externally, 0 reviewed independently), version or commit, the reproducible test, the result with date and machine, and independent verification (none yet for every row). Evidence link in the homepage nav and the generated pages' nav.
docs/benchmarks/proving-e2e.md: replaces the 20-second shard gate with three fixed workloads, job-received-to-accepted-proof latency, cost per proof, the eligible card list with mining and proving reported separately, the verbatim acceptance standard and the three-unrelated-operator protocol.
docs/plans/funding.md: cost, what is funded (founder's means, the client's 1% fee once there is mining), what waits on revenue, what pauses.
docs/analysis/security-budget.md: emission through six halvings at three price inputs, miners and provers separate from burns, the USD 1M floor and the year it is crossed.
docs/design/payment-routes.md: Mermaid flowchart and table of every flow, with operator, app, team and protocol revenue labelled.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
proving/igneum-prove: core (port of igneum-exec at fb33069 as the block statement), program (SP1 v6.8.1 guest),
host (execute, core, compressed; ProofSystem trait with the stub and the SP1 implementation), export (cuts a block
out of igneum_exportSegments and checks every state root against the node's). Fixtures block-78-increment and
block-56-transfers from the 3-node simnet. proving/windows-wsl2: SETUP-PROVER.bat, setup-wsl.sh, PROVE-BLOCK.bat,
make-package.sh. docs/plans/proving-v0.md: the devnet v4 shard plan, what tonight's proof shows and does not, the
morning acceptance line. Mac CPU baseline (block 78: 626 k cycles, core 22.0 s and 7.3 MB, compressed 55.7 s and
1.27 MB, both verified) appended to docs/bench-log.md, left uncommitted because that file carries another agent's
pending changes.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
igneum-seed-1 (Hetzner cx23, fsn1, 188.245.5.161:26611): built on the VM in 1,530 s, synced to the live devnet
(12,204 blocks, same sink as the live node) through a non-mining relay igneumd on the Mac (the live node's addPeer
RPC is refused in safe mode); the live node and the Windows PC learned the seed's address by peer exchange and dialled
it. seeds.txt written. Hetzner prices corrected to USD (pricing API currency) in the plans, READMEs and scripts;
current-generation types per location (cx23 EU, cpx22 sin, cpx21 US) in the cloud-devnet config.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
tools/harness runs the standard consensus-attack catalogue against a private
test network of our own igneumd nodes (127.0.0.1 ports 27200+, /tmp/igneum-harness,
never the live devnet or the PC node), with a pass criterion per scenario from the
spec and a measured result each. Built on the node fork's own crates
(igneum-harness-sim on kaspa_utils::sim as simpa does; igneum-p2p-probe for the
wire). Scenarios: 1 withholding, 2 timestamp edges and drift, 3 partition and heal,
4 eclipse, 5 malformed and boundary inputs on every p2p and RPC surface, 6 resource
exhaustion, 7 fast-miner flood. Finality and difficulty-controller scenarios are
stubs with their criteria written.
bench-log: one dated entry, a row per scenario (criterion, measured, pass or fail).
First run: 19 of 20 measured rows pass. Findings recorded in the entry: scenario 5
reproduces ledger M15 on HEAD (bogus past-day or DAA headers build a 256 MiB cache
before rejection; the r3-fixes branch removes it); scenario 1 at 45% hash with
burst withholding shows a selfish-mining blue-share gain (50.7% of blues), the one
failing row.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/design/execution-layer.md section 10: what the execution-layer branch implements
(D1 to D10, RPC, differential), the devnet rules fixed there, what is missing, the merge
plan with the finality branch. docs/bench-log.md: the 3-node simnet run with numbers.
tools/evm-smoke: viem 2.57 smoke test (fund, 50 transfers, duplicates in parallel blocks,
contract deploy and call, state roots across nodes, export for igneum-exec-diff) and the
solc build of DeveloperRegistry and the Counter test contract.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
infra/cloud-devnet: hcloud (doctl variant) create, builder-VM provision from a git-archive source tarball,
systemd units for igneumd --devnet-suffix with a sparse --addpeer mesh and a CPU trickle miner per node,
stdlib wRPC client, experiments (latency, partition, hop, collect, observer hookup), README with the command
sequence and the Hetzner API prices of 3 Oct 2026.
infra/gpu-bench: RunPod image recipes (CUDA 12.8, ROCm), bundle, run.sh (vectors gate, 10-min raw, sweep,
inline shortcut ratio, nvcc/NVRTC/OpenCL recompile timings, results row, intake upload), bench-log template.
infra/seed-nodes: create-seed (persistent IPv4, firewall), provision on the VM, health check, addPeer from the
Mac over grpcurl, seeds.txt; igneum-seed-1 created at 188.245.5.161 (Hetzner cx23, fsn1).
docs/plans/cloud-devnet.md and docs/plans/seed-nodes.md.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- docs/fork-divergence.md: "Finality v2" table (every file, risk, merge note), decisions
- docs/spec/03-finality.md: section 3.10 implementation notes, clause by clause
- docs/bench-log.md: test-network results (72 of 72 steady locks, median 0.80 s; equivocation
strip; partition: 0 locks at 39.6% of total with the floor binding, heal in 30 s), follower
- tools/observer: live_checkpoints table, FinalityLock subscription, "checkpoint N locked" events
- site: /api/live adds checkpoints and locked/final flags; /live draws the lock ring and final line
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Pool protocol after Stratum V2 job declaration: member-checked or member-built templates, JSON over TLS, shares at target64 << s on the 32-lane unit, one vote key per operator held by the member, votes relayed and carried by the pool with a chain-only drop test. Light client: trust table, checkpoint-mode bytes per day, pinned seed list, the read-only node API, the homepage card's steps. Phone app: wallet, miner monitor, node card, store rules, build plan on journey.json.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
fork-divergence.md: a section for branch r3-fixes (the PoW-before-validation
reorder, the cache-build cap and the per-peer guard), the merge-overlap note for
the finality branch, and the updated "PoW before or after GHOSTDAG" and "Day
seed" open decisions. bench-log.md: the before-and-after attack numbers (50
bogus headers build 50 caches in 10.6 s before, 0 and all rejected in 14 ms
after), one cache builds in about 0.2 s, and the M16 Mac inline-dataset note.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
proto-cuda/windows-node/: START-NODE.bat and start-node.ps1 (igneumd
--devnet with --addpeer to the Mac, status line every 30 s through
igneum-miner watch, keep-awake, random-delay restart, Ctrl+C),
BUILD-NODE.bat and build-node.ps1 (builds on the PC from the src.zip
snapshot; winget for Rustup, LLVM and protoc; MSVC default toolset),
ALLOW-FIREWALL.bat and allow-firewall.ps1, README.txt, cross-build.sh
(the mingw-w64 recipe that built igneumd.exe in 8 min 25 s; the exe
ships with the three mingw runtime DLLs) and make-package.sh.
WINDOWS-MINER.md gains "Run your own node"; bench-log records the
cross-compile and the two-peer sync test on the Mac (ports 27000 and
27010, live node untouched). The mining launcher's NODE_HOST=auto
edit stays uncommitted for the agent that owns start-mining.ps1.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Litepaper Finality: the reorg bound users rely on is the 12-hour finality depth in median time; Kaspa's one-hour merge depth is a merge limit, not a reorganisation bound; first-month sentence and "does not claim" item 4 say the same. Litepaper Speed: emission per block on a schedule keyed to difficulty-adjusted time, reds in the window paid, coins track blocks within the controller's accuracy.
Design 5.5 and D12: the native-execution veto is relative to the carrying block's own selected-parent chain; two-node reorg test added to 8.5.
Ledger P11, F15, E10 statuses and fixes rows 79, 82, 84 updated.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The four public sentences (F18, P13, E11, L7 in site/litepaper.html and site/index.html) were swept into the concurrent commit cd604db; this commit carries the rest.
Spec 02: finality depth (43,200 DAA s, 12 h of median time) named as the reorg bound, merge depth as a merge limit only, simnet reorg test for gate 2; emission follows the code (365.25-day year, 63,115,200-s halving, 31.688 IGN per DAA second, reds inside the DAA window paid to the merger, E paid per block so coins are blocks times E).
Spec 03: W2 and Q1 denominated in past-median time, weight as a share of each 60-s bucket; W6 keys are free, weight is the only Sybil-resistant quantity; 3.8 and 3.9 point exchanges at the finality depth.
Spec 06: O-3.14, the finality simulation with the DAA in the loop under a pulsed rental (gate 3).
Spec 07: shard sortition draws by weight (blue blocks drawn uniformly from the window); proof-record validity is relative to the carrying block's own selected-parent chain; 1-key-versus-1,000-keys test.
Spec 08: release-key chain, rotation signed by the current key, revocation signed by the previous key, both published in a block; policy before the client ships.
Ledger: status lines for F18, P13, E11, L7, P11, F17, F14, M14, F15, E9, E10, G9; P8 cross-reference. Fixes: section 2.2 rows 75 to 88, with M15, M20 and P12 marked code, owner consensus engineer.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Seven reviewers (Kaspa core, RandomX author, Ethereum client, GPU farm,
chip designer, exchange listing, regulator's analyst), three attacks each
against docs/spec, the execution design, the fork code and tonight's devnet.
0 fatal, 18 serious, 8 minor. New ledger entries M14 to M21, F14 to F18,
P11 to P15, E9 to E11, C13, L7, L8, G9, G10, X12; one cross-reference on P8.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/provenance.md: table of every component Igneum uses (origin, licence, what changed, why, how measured), what is new, what to adopt from upstream, own-code licence pending the project lead's decision. Licences verified on disk: rusty-kaspa ISC, chiavdf Apache-2.0, igneum-pow MIT, blake2b_simd MIT, blake3 CC0 or Apache-2.0, sha2 MIT or Apache-2.0, secp256k1 CC0, keccak Apache-2.0 or MIT. RandomX, SP1, revm, blst, ProgPoW, LWMA, Monero, GMP, sha3: approximate, not cloned.
site: litepaper gains the Built on the shoulders section and nav entry; index gains the two-line mention and footer link near the RandomX comparison; the block rate reads one block a second at launch, rising, where it read as permanent (litepaper diagram, index live section).
tools/upstream: README with the exact merge commands, expected conflict files from fork-divergence, the test list and the consensus-review rule; sync-upstream.sh fetches and opens the merge on a branch without committing. Not run against the fork.
docs/commercial/prover-customer-brief.md: one-page brief for a first proving customer at testnet, timeline from journey.json, risks, 10 candidates labelled approximate.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Records the 3 Oct 2026 rename pass in vendor/igneum-node (binary, process
name, user agent, data and log paths, env vars, address prefixes, DNS
seeders, default build set) and what stays Kaspa-named internally. The
Windows miner guide and the observer README now start igneumd.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
proto-metal --serve compiles igneum_hash_bound at runtime and mines jobs from stdin (init words in buffer 3, program
and dataset cached per seed); proto-cuda and proto-opencl host serve modes from the pack's kernel_bound.cu / .cl with a
seed guard; --vendor device filter for OpenCL. windows-miner/: START-MINING.bat + start-mining.ps1 (GPU and tool
detection, pack export, cached builds, MINERS identities per vendor, status every 30 s, uploads every 60 s, rebuild on
seed change, Ctrl+C summary), README.txt, make-package.sh (igneum-mine-test.zip with a cross-compiled miner).
docs: fork-divergence devnet v1, bench-log entry with the CPU, Metal and overnight numbers.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Preliminary clearance across EUIPO, UK IPO, USPTO, WIPO and the UAE for
IGNEUM in classes 9, 36 and 42, with the hit table, crypto-name
collisions, handles, and the ADGM DLT Foundation filing strategy.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- 03-finality: Q2 participation from every vote seen in blocks (votes are
block payload); 3.3.2 the eclipse case closed by the 56.7% floor with the
sim v2 F2 numbers; 3.4.1 why aggregators cannot grind participation.
- 07-execution (new): EVM semantics on the DAG, shard sortition (8 provers,
10 s, then open, no shard bond), bridges (none official, no bridged
stablecoins at genesis, proof bridge with the consensus proof in phase two).
- 08-client-security (new): reproducible builds, release key in genesis and
in hardware, no silent updates, consensus only by 90% signalling, notarised
builds, official sources with the hash, seed confirmed before mining,
hardware wallet, the permanent seed line.
- 00, 02, 05, README, 06: cross-references, O-2.8 removed, O-3.3, O-3.7,
O-5.1, O-5.2, O-5.6 narrowed, O-8.1 added, counts kept at 60.
- Ledger: eight Status lines, status table, count table, overclaims 38, 40, 71.
- FUD fixes: rows 23, 26, 38, 62, 64, 66, 67, 70, 72, section 3 and 4.
- Site: bridge and stablecoin sentences no longer launch features; the seed
line wherever the app appears.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Every open, conceded and answered-with-remainder entry of docs/fud-ledger.md
as one row: fix, owner, timing, and whether it exposes the team. The 78
overclaims checked against the live site text. What the 3 October decisions
close. The pre-public-repo scrub, in order, including the history rewrite.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Decision of 3 October 2026: no fee to any team, foundation or fund.
Priority fee now splits 80% to the miners and provers of the block and
20% to the application called, attributed per call frame as before.
No burn of the tip; the base fee is still burned in full on both gas
dimensions. External proving jobs pay 90% to the provers who delivered
and burn 10%. The 60% signalling mechanism stays as a general tool for
parameters that genesis rules leave to miners; upgrades stay at 90%.
Ledger E4 closed by removal; E3, E5, G4, G5, G8, L1 and overclaims
45, 46 and 53 rewritten to the new rule.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The ledger stays in the repository as an internal working document while it carries open items. /ledger now redirects to the homepage. Promotion-rules entry rewritten without any founder location.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/spec/01 to 06 and the README index, alongside the existing overview. Section 1 is the normative
definition of the lottery hash with test vectors copied from the igneum-genesis-mh pack (cache fingerprint
48c4f5bf24166b2e, 96 hashes, mixer constants) and every prototype value marked with the measurement that
fixes it at gate 1. Sections 2 to 5 are the design as decided on 3 October 2026, labelled Designed. Section 6
lists 60 open items with the experiment or decision that closes each and its gate.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/fork-divergence.md: every rusty-kaspa file the fork changed (file, what,
why, risk, upstream-merge note), the decisions left open (8 vs 18 decimals,
temporary epoch seed, day seed, lane-to-target mapping, pool payee, depth
bounds, PoW after GHOSTDAG) and the per-second subsidy table.
docs/bench-log.md: 3 October 2026 entry for the 3-node igneum-devnet run at
0.84 blocks/s with the 80/20 coinbase and vote_key_hash verified on all nodes.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The chain animation scales its proving tempo to the canvas width, so a block is mined, proven and locked before it leaves a phone screen. Opening frame seeds proven and locked blocks on narrow canvases too. Repository links point at github.com/igneum-network/igneum.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Exporter writes kernel.cl next to kernel.cu (same instruction list; memory-hard core emitted in a third, OpenCL C
dialect with the same literals as memhard.h). Pack headers are now C99-safe so a plain C host can include them.
proto-opencl/host.c: C99 + OpenCL 1.2 API, device list, runtime build, cache fill and FNV check, dataset build and
self-test, 3 vector warps standalone and in batch, bench and sweep as host.cu, whole-batch fingerprint. The 32-lane
exchange is sub_group_shuffle_xor only when the queried sub-group size for a 32-item work-group is exactly 32;
otherwise a local-memory exchange with one barrier per exchange, so wave64 hardware cannot change the hash
(WAVEFRONT.md). build.sh (macOS, Linux), build.bat (MSVC), README with the exact AMD-rig commands.
Proven without AMD silicon: Apple OpenCL 1.2 on the M5 Max 96/96 on all three packs (45.0 Mhash/s at 1 GiB, Apple
number, not AMD); pocl 7.2 CPU device 96/96 on both exchange paths including the real sub_group_shuffle_xor text;
CPU emulator 7 configurations incl. 64-wide sub-groups, identical fingerprint f99fb375b3abeaf5 everywhere.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Class group (1024-bit prime discriminant from the checkpoint hash, chiavdf
construction, NUDUPL/NUCOMP/Lehmer xgcd ported from vendor/chiavdf) and an
RSA-2048 trusted-setup stand-in for timing. eval, block prover, verify,
epoch_seed/verify_epoch_seed, grinding model, README with measurements and
the parameter recommendation, bench-log entry.
M5 Max: class 163k sq/s (T 98 M for 10 min, 588 M for 1 h), verify 4.5 ms,
proof 516 bytes; full 10-min runs for both groups; grinding gain for a 30%
miner +3.62 blocks/epoch with no delay, 0 with it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
proto-metal: default dataset is now the memory-hard construction (MEMHARD.md), --closed-form keeps the original.
Cache fill 2 ms GPU / 185 ms one CPU core; dataset build 20.6 ms; GPU cache == CPU cache on all 2^26 words.
Shortcut ratio: inline kernel 111x faster than honest (closed form) to 4.8x slower (memory-hard), 1 GiB.
CPU verify 0.63 to 0.80 ms per warp at 104 loads, 1.21 ms at 144 loads (4,608 items): 10 ms gate met.
Levers --load-weight and --wide-frac implemented and measured, both off; default generator unchanged.
Fuzz 200/200, edge, determinism, memcheck, stats re-run on the new dataset, all PASS.
proto-cuda: host.cu handles both dataset modes; new pack igneum-genesis-mh with memhard.h; clang emulation PASS
including the three-way cache check. Old packs unchanged; closed-form export is byte-identical to them.
docs/bench-log.md: dated summary.
Note: a concurrent session running git commit -a swept earlier states of these files into its site commits
(7b28d5e through d6539fa); this commit carries the remainder.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
