dimension version 3 — the root chain's genesis version, deferred hashing, both burials committed

Date: 2026-10-16. Scope: bbg feat/dimension-v3 — the 0.3 line (chore/coordinated-release-20260916, ccda7a2) + #40 (launch row 43, incremental commitments) + #41/#42 (dimension version 2) + the per-neuron nonce of feat/chain-dims (97c6272) + this branch. Consumer: soft3 feat/bbg-nonce-on (the chain kernel and soft3 genesis build). Trackers: cyber/launch.md rows 43 and 14. Machine: M4 Max, 16 cores, 48 GB, release builds; the machine was shared with other agents' test runs (load average 30–70), so single-thread costs are reported as user CPU time beside wall time.

what changed

1. NeuronRecord.nonce: u64 — the chain's replay counter under the root (soft3 specs/chain.md §2.5, §3.2); fourth neuron value in the trie leaf, the flat table (width 8 limbs) and the native record (32 bytes).

2. DIMENSION_VERSION = 3 names the rule set (specs/state.md § dimension versions) and enters every trie leaf IV: [LEAF_TAG, |v|, 3, 0]. The flat table layout keeps its own word, TABLE_VERSION = 2 — the zheng state-evidence contract, record-agnostic — so zheng release/0.4 reads bbg certificates unchanged.

3. Deferred hashing: a mutation marks the key (Keyed, AxonIndex) and the trie path (Trie); the next root read hashes each touched leaf once and each dirty branch once, bottom-up. Above 2^14 touched records the dimensions hash on parallel threads (a bulk load). Roots unchanged.

4. AxonIndex: degree histogram (max degree without a scan), truncate (the kernel's revert of a block's appends), the list mirrors the set.

5. Vendored fjall builds warning-free.

6. Tests are hash-agnostic: every expected root is recomputed by an independent reference of specs/state.md over hemera (no hand-pinned digest), so the vectors follow hemera 0.4 without re-pinning.

7. (amended 2026-10-17, rule 7) the signals dimension is keyed by signal_key(neuron, step) = H(neuron ‖ step) and SignalRecord carries step — two neurons' step-0 signals no longer share a record (launch row 9). Version 3 kept; no blessed root moves. Evidence: signals-key-2026-10-17.md.

Amended 2026-10-17: version 3 is no longer the root chain's genesis version. Dimension version 4 (rules 1–7 unchanged, rules 8–10 positions; soft3 specs/chain-economy.md §3) replaces it before genesis froze, and the version word in every trie leaf moves every root recorded below. Evidence: positions-v4-2026-10-17.md.

why not chunked Brakedown commitments

The alternative considered for the size cap: split each flat dimension into Brakedown-committed chunks of ≤ 2^20 fields, hemera-hash the chunk roots into the dimension leaf, recommit only touched chunks. Brakedown commit cost, lens release/0.2, one thread:

bostrom's adjacency dimensions are 2^25 fields each (26–27 chunks). A sorted flat layout shifts every chunk after an insertion point; with key-bucketed chunks a block of 100 random links still touches most chunks of each adjacency dimension: ≈ 50 chunk commits × 1.39 s ≈ 70 s per block against a 30 s block target. The trie costs ≈ 0.3–0.7 ms per link, has no cap, and its worst case under adversarial keys is bounded (depth ≤ 256; chunks under adversarial clustering are not). Chunks would keep Brakedown cell openings, but the zheng look argument cannot open a chunked leaf without a circuit change either; when zheng opens state it opens trie paths (compress4, the primitive the root fold already replays). One hemera compress4 costs 2.7 µs here.

row 43 — commit cost, before and after

rs/examples/commit_cost.rs N 100 [v2|v1] (synthetic one-link signals with conviction, 1,000 neurons, Zipf-like targets, root at every block edge). "v1" is the flat derivation that was the default before #40; "trie, eager" is #40 (c95924b, hashes every path on every write); "trie, deferred" is this branch (cad9f40+). Same roots in both trie columns (bb22fc74… at 10k, 4ad5342e… at 100k).

Per-link cost grows with log n only (depth of the touched paths), never with n: v1 is linear per block, so quadratic per chain. Deferred hashing pays per distinct record per block and shares the upper trie levels: a single 10,000-link block costs 134 µs/link, one link per block 2.2 ms/link (three root reads per block: before time[h], after, checkpoint).

Root bit-exactness: dimension version 3 moves every root on purpose (nonce, versioned leaf IV). Within version 3 the eager (#40 design) and deferred implementations give identical roots on every vector, every property test and both burials (pussy: identical genesis root 93e4660f… from both binaries).

recovery

• bbg recovery_replay_reproduces_every_stored_root: 60 random blocks (conviction-free and paying links, nonces bumped, zero balances dropped), replayed from scratch — every stored block root time[h] reproduced, and a from-scratch rebuild of the final records gives the same root.
• bbg bulk_and_incremental_commitment_reads_agree: one root read after 12,000 writes (parallel flush) == a read every 97 writes == a rebuild.
• soft3 reopen_replays_bodies_and_verifies_roots (chain kernel recovery: replay every main-chain body on genesis and check every header's state_root) and the fork/reorg suite pass on this bbg.

property tests

• trie: boundary keys (all-zero, all-one, first-bit and last-bit splits, the 256-deep single-bit chain, u64 edges) under 1,500 random set/remove steps vs from-scratch, with inclusion/non-inclusion proofs for every key every 25 steps; random batches vs from-scratch; the dimension version is in every leaf (recomputed from compress4).
• AxonIndex: 2,000 random add / remove_axon / truncate / remove / insert vs a model, commitment == a fresh build, max degree == the model's.
• state: incremental == from-scratch after random batches, pruning and replay; nonce under the root under v1 and v2 (including the get_mut guard); both root versions == the reference derivation (empty, one link, every dimension, before/after a block).

genesis, both burials (soft3 genesis build)

See soft3 audit/genesis-dimension-v3-2026-10-16.md for the full table.

Before (dimension version 2, eager trie, #41's run): bostrom 1,790 s and 8.57 GB; before that (flat v1) the bostrom root was blocked by the 2^20 cap.

what this does not do

• zheng opens no trie paths yet: look proofs and state certificates pin RootVersion::V1 on small states.
• The tries live in memory beside the maps; a node rebuilds them on load (the deferred flush makes that one hash per node).
• hemera 0.4 (profile #15) will move every root once more; no bbg code pins a digest, so only consumers' blessed vectors re-bless.