soft3 on hemera 0.4 — the phase-1 root chain re-derived under profile #15

Date: 2026-10-16. Sweep part 1b of hemera 0.4 (profile #15, merged on hemera release/0.4 7237a23; receipt hemera/audit/profile-0.4-2026-10.md §5 names what soft3 must move). Branch soft3 feat/hemera-0.4 on origin/main 2d55fdb. Contracts: specs/chain.md, specs/chain-conformance.md, specs/genesis.md.

Stand ~/cyber/.stands/S1B, detached worktrees: soft3 origin/main, hemera origin/release/0.4 7237a23, bbg origin/master eb152fa, zheng origin/release/0.4 09308d7, lens origin/release/0.2 e11eaa8, nox / joy / trident origin/release/0.4, cybergraph feat/durable-applications 0eae7ae, mudra refactor/spell 4bfe3e8, foculus feat/device-profile-choice b28f0a2, honeycrisp origin/main (lens gpu/'s aruminium path), tade / strata / tru / tok / radio / neuron / inf / file defaults. Machine: M4 Max, 48 GB, shared with other agents' builds throughout (1-minute load 20–90). Build: cargo build --release --no-default-features — the CPU miner only; the GPU miner follows lens (agent S1a, lens feat/hemera-0.4, not yet on release/0.2 when this ran).

1. the dependency

• crate/Cargo.toml, conformance/rs/Cargo.toml: cyber-hemera 0.3 → 0.4 (path kept); both lockfiles move the path entry to 0.4.0. radio keeps its own git-pinned cyber-hemera 0.3.1 (d5f0a09) inside the wire crates — radio's sweep, not soft3's; the chain's topic ids are computed by soft3's hemera.
• release/phase1.toml: hemera pin main 23f3bbc → release/0.4 7237a23afad38913a143da48010f3912f50d7684, updated = "2026-10-16".
• conformance/rs/Cargo.toml also moves bbg 0.2 → 0.3, nox 0.2 → 0.3, zheng 0.3 → 0.4, lens 0.1.3 → 0.2: the committed requirements did not resolve against any of the pinned siblings.
• Every sibling still requires cyber-hemera = "0.3" with its path dep (bbg rs/Cargo.toml, cybergraph, mudra, foculus, lens workspace, nox, joy, trident, tru, tok, zheng). Cargo refuses a 0.4.0 path crate for a ^0.3 requirement, so each was bumped to "0.4" in the stand only (uncommitted). bbg master does not "follow hemera automatically" until its one-line bump lands; the same line is owed by every repo above (hemera receipt §5, first paragraph). Nothing in soft3's dependency graph uses the removed API (Goldilocks::inv, PartialSbox, Profile::new, RoundVisitor); the .inv() calls in foculus, nox, lens and strata are on nebu's field, not hemera's.

2. the vectors (specs/chain-conformance.md)

Every pinned digest in crate/src/chain/{sigs,header,pow,tree,genesis, net/codec}.rs, genesis/tests.rs, wallet/tests.rs, mine/{mod,bench, target}.rs and §1, §5, §6, §7, §12 of the spec was re-derived through the chain's own code (a throwaway printer linking the crate; the tests then assert the same values). The definitions did not move.

The full §12 table (owner, rhos, commitments, nullifier, roots, spend id, vk set, statement id) is in the spec; chain.snap carries their fingerprints.

the mined header, re-mined

The §5 PoW-valid header (bits = 0x1e010000, time = 1793885850, the test header's other fields) was re-mined, not replaced: a linear nonce search from 0 through Header::id (= hemera::hash over the 168 bytes) and valid_pow — the kernel's own path, no miner backend.

nonce    934636 = ec420e0000000000 0000000000000000        (0.3: 40165933)
block_id a3dc0cdb7c69e1d4986c388effe1612613b1b71f2c2ca49084bc0730da000000

934,637 hashes; one thread 10.4 s = 90.3 kH/s on one core (0.3: 64.4 kH/s, 623 s for its 40.2 M hashes) — 1.40× for the sponge over 168 bytes, against 2.2× for the bare permutation (hemera receipt §3: the sponge's four permutations are latency-bound like the single row). Twelve threads found the same nonce (least index, 1.52 M hashes, 2.2 s). Byte 28 is 0xda, so the harness's rejection at bits = 0x1d100000 (needs byte 28 below 0x10) holds; nonce 934,635 fails. chain.md §1.2, §1.3 and §8 carry the new rate (order 0.9 MH/s per ten cores, floor block ≈ 12 s on one core); the mine bench vector check and the backend-agreement test search from base 934,000 for index 636.

The CPU miner (permute_batch, 16 lanes) measured 91–104 kH/s per thread inside the fleet below on the loaded machine; hemera's receipt measures 0.4 batched at 0.6× of 0.3 batched, so the CPU miner is not faster on 0.4 even though single hashing is. The miner's sponge view (mine/job.rs) is unchanged: same rate, same limbs, same length lane.

3. genesis

soft3 genesis build --chain <c> --snapshot ~/cyber/<bootloader>/snapshot --out ~/cyber/.cache/genesis/<c>.tade (cache ~/cyber/.cache/snapshots/<c>, pinned inputs by sha256; bbg dimension version 3). derive.rs re-derives every particle (hemera(cid), the black hole of launch row 20 — 74,514 CIDs have no bytes on the cybernode) and every owner id.

The file sizes, counts and sums are byte-for-byte the 0.3 shape; only the ids moved. Previous run (0.3, same day): audit/genesis-dimension-v3-2026-10-16.md (bostrom 204–268 s, 7.5 GB) — this run was slower on a machine at load 70–90, not from 0.4 (user time 110 s against 189 s).

4. conformance

• soft3 conformance bless chain --genesis ~/cyber/.cache/genesis/pussy.tade: 30 of 30 entries moved (every entry is a hemera value); soft3 conformance check chain green against the rebuilt file; conformance/chain-genesis.toml rewritten with both chains' values and the hemera pin.
• cargo test --release --no-default-features --features chain-conformance --locked: 26 of 215 lib tests failed on the old vectors before the sweep (every one a pinned digest, none a logic failure); after it 215 lib + 12 wallet + mine e2e + adversary pass (sections 1–9, 11, 12; §10 through the cached pussy file). The chain tests do not depend on zheng proofs: no zheng reference in crate/.
• The stack conformance harness (conformance/rs, snapshots mechanism.snap / encoding.snap / manifest.snap / provenance.toml, surfaces/hemera.rs) could not be re-blessed: it links zheng, whose release/0.4 head 09308d7 does not compile against hemera 0.4 (RoundVisitor::partial_round takes [x², x³] now — the gadget in rs/src/execution/hash.rs, hemera receipt §5) once its legacy surface is on, and the harness's surfaces/zheng.rs uses only that retired API (commit/open/verify_eval/row_hash, gated behind legacy in 0.4). Waits on S2 (zheng feat/hemera-0.4), then on a zheng-surface rewrite. nox release/0.4's hash pattern is still the 24-round 0.3 trace (rs/patterns/hash.rs); its surface entries move again with nox's sweep. The hemera surface's doc line and conformance/specs/README.md name 0.4; the .snap files are untouched (still the 2026-10-10 bless).

5. the fleet

nu scripts/chain-fleet.nu --threads 1 --keep --http-base 17980 --radio-base 17990 (the flags are new: the default range 17780/17790 was held by another session's SSH tunnel and pussy-rc node, and the script only kills what it spawned). Run 1 (20 s window, load 70–90, cargo builds running): peers connected, 2 blocks in 20 s — RED on the ≥ 3 check: 2^20 expected hashes per floor block at 91–104 kH/s per thread is ≈ 10 s per miner, 5 s per block for two miners, and the window was short on a loaded machine (the script's comment "a block per 4 s" is a 0.3-batched figure). Run 2 (40 s window, load still 70–90): GREEN — peers connected (2 each); heights agree at 5 across 3 nodes, block 1 aea953a4… identical everywhere; the late joiner IBDs to height 5 by ranges; a tx submitted to node 1 is mined at height 8 by node 2's miner and bob holds 5 on node 3; a fork of depth 3 from height 8 (main 11, fork 13) reorgs every node to the heavier chain; node 3 killed at height 13 restarts from disk, replays, syncs to 17, conserved. Log: ~/cyber/.stands/S1B/fleet-2.log (stand deleted after the sweep; the run is reproducible with the command below).

6. what waits

• S1a (lens feat/hemera-0.4 on release/0.2): the Metal / WebGPU miner (gpu/: ginv, RC[144], the shared inversion). This branch builds --no-default-features; every_backend_agrees_and_verifies and mine bench run the CPU backend only. When lens lands, the default features build again and the GPU backends must find index 636 from base 934,000.
• S2 (zheng feat/hemera-0.4): the stack conformance re-bless (§4), the zheng proof fixtures under conformance/fixtures/, and surfaces/zheng.rs.
• sibling bumps (cyber-hemera = "0.4"): bbg, cybergraph, mudra, foculus, lens, nox, joy, trident, tru, tok, zheng — one line each; until bbg's lands, soft3 main does not resolve against hemera release/0.4 without a stand-local edit.
• cyber/launch.md row 14 carries the new ids (cyber PR).

reproduce

cargo build --release --no-default-features                      # crate/
cargo test --release --no-default-features --features chain-conformance --locked
soft3 genesis build --chain pussy   --snapshot ~/cyber/space-pussy/snapshot --out ~/cyber/.cache/genesis/pussy.tade
soft3 genesis build --chain bostrom --snapshot ~/cyber/bostrom/snapshot     --out ~/cyber/.cache/genesis/bostrom.tade
soft3 conformance check chain --genesis ~/cyber/.cache/genesis/pussy.tade
nu scripts/chain-fleet.nu --threads 1 --secs 40 --http-base 17980 --radio-base 17990