commit cost — before and after incremental dimension commitments

Date: 2026-10-15. Scope: bbg 441e3c9 (origin/master) → branch fix/launch-43-incremental-commit. Tracker: cyber/launch.md row 43 (blocks row 14, the bostrom genesis replay). Prior finding: cybergraph audit/import-commit-cost-2026-09-27.md — every accepted operation recommitted every dimension from its full record set, so a chain's cost was quadratic in its length.

what changed

• every committed dimension is a Keyed map (KeyedSet for N, AxonIndex for adjacency) that keeps a canonical keyed trie (rs/src/trie.rs) current on every mutation and caches its v1 Brakedown commitment until it mutates;
• RootVersion::V1 (default) folds the Brakedown leaves — bit-exact with the previous root, only dimensions touched since the last read are recommitted;
• RootVersion::V2 folds the trie roots — a block costs O(changes · log n);
• max_degree is maintained by the adjacency index, not scanned per root.

Spec: specs/state.md § BBG root, § root versions, § dimension commitment v2.

method

rs/examples/commit_cost.rs N B [v1|v2]: N synthetic one-link signals from 1,000 neurons, sources uniform over N/2 particles, targets Zipf-like over N/4 (hubs), one block per B = 100 links, root materialized at every block edge. rs/examples/replay_pussy.rs CSV [v1|v2]: the pussy snapshot (~/.cache/snapshots/pussy/cyberlinks_indexed.csv.gz, 29,112 cyberlinks, 114 neurons), one block per source height (17,696 blocks). Release build, one thread, M4 Max; peak memory = /usr/bin/time -l "peak memory footprint". Wall time is the whole run including state construction.

results

synthetic, 100 links per block

Per-10k-link cost along the run (the curve, not the average):

v1 grows linearly with the state (each 10k adds ≈ 8 s per 10k), so the chain is quadratic; v2 grows with log n only (the trie depth), ≈ 10 % per decade.

(v1 per-10k numbers are from the 100k run, which shared the machine with builds for its first minutes; the standalone 10k run is the 39.6 s above.)

pussy snapshot, one block per height

cost model after

Per touched record: one leaf chain (1 + ⌈fields/4⌉ compressions) plus two compressions per branch on its path (depth ≈ log₂ n). A one-link signal touches particles ×2, axons_out (outer leaf + inner set), axons_in (same), neurons ×1, balances ×2 ≈ 8 paths ≈ 300 hemera permutations ≈ 0.6–0.9 ms on one M4 Max core. Projection for the bostrom replay (2,949,732 links): ≈ 45 minutes single-threaded; the spacepussy-test log (39,060 signals): under one minute. Both inside the row-43 bounds (one day / ten minutes).

Headroom, in order: hash the branch with one permutation over [tag, bit, 0, 0 ‖ l ‖ r ‖ 0⁴] instead of two compress4 (−45 % per path; a different in-circuit row shape than the root fold, so it needs zheng's word); batch the neurons[ν].focus updates of a multi-link signal into one leaf write; parallelize dimensions.

root

v1 roots are unchanged: tests/root_vectors.rs pins five roots computed on 441e3c9 before the change (empty, one link, one link finalized, every dimension populated, same finalized) and they hold. All 64 unit tests, the look-argument e2e and the wire tests pass without modification.

v2 roots differ from v1 by construction (different leaf derivation) and are pinned too. Properties checked: incremental root == from-scratch root of the final record set after 40 random batches with double-spend rejections and block finalization; after pruning and reinsertion; replay reproduces every per-block root; trie paths prove inclusion and non-inclusion against the dimension leaf; v1 ↔ v2 is a view switch on the same records.

No conformance snapshots exist to bless: soft3/conformance is a scaffold (no .snap files in soft3, bbg or cybergraph on 2026-10-15).

what this does not do

• It does not flip the default. The zheng look argument binds LookOpening.commitment (a Brakedown cell opening) to leaves.dims[ns]; under v2 that leaf is a trie root and the binding fails by design. v2 becomes the default when zheng opens trie paths (leaf chain + branch chain, the same compress4 rows build_root_steps already emits). Until then a node opts in with Bbg::with_root_version(RootVersion::V2).
• It does not change cybergraph. prepare_native_batch (cybergraph feat/durable-applications, bbg feat/atomic-application-storage) may keep calling refresh_root per accepted operation: under v2 that call costs the fold plus the operation's own changes.
• Storage is untouched: the tries live in memory next to the maps (v2 1M links: see peak memory above). A node that loads state from storage rebuilds them by inserting through the wrappers, O(n log n) once.