diff --git a/CHANGELOG.md b/CHANGELOG.md index 5e1473ca..74985a02 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -184,6 +184,87 @@ Docs (issue #119, AGG-E2E-2; no code change): the build that cuts, which constraint set 10 moves from `792a66d4…` to `a3d5c38b…`, issue #166). `docs/deploy.md`, `docs/compute-optimization.md` and `README.md` follow. +Node, network and deploy (monorepo audit-v14 issues, merged 2026-11-18/19; no consensus change — +what each needs before it takes effect is in its PR): + +- **A committed systemd drop-in turns the public RPC listener on** (VK-2; deploy only, no + consensus or network change): `deploy/systemd/rand-node.service.d/60-public-rpc.conf` sets + `RAND_NODE_PUBLIC_RPC=127.0.0.1:8546`, which `--public-rpc` now also reads; `DROPIN=1 + deploy/enable-public-rpc.sh` installs it. New `deploy/check-public-rpc.sh` checks a listener, + tunnel end or host — the viewing-key methods, `rand_mint`, `rand_getPeers` and batches must be + refused by the node's public listener — and its output is the deploy record's evidence. The live + hop's check is still to be run and stored. (#186) +- **`deploy/sign-release.sh` no longer erases existing signatures** (PROC-5; release tooling + only): without a flag it refuses where a `SHA256SUMS.sig*` exists; `--append` adds a second + holder's signature beside the first, refusing a duplicate signer or a signature set that is not + over this `SHA256SUMS`; `--replace` starts over. Self-tests for the two- and three-holder flow + and for k-of-n verification run in CI. Still one release key listed and a threshold of 1: the + second key and the raise to 2 are the owner's. (#188) +- **A node stems around a successor that acknowledged a stem and went silent before it publishes** + (PRIV-604; network behaviour local to the node, no wire or consensus change): the first time a + held transaction's embargo runs out the node sends it to a successor that has not had it, under + a fresh embargo, and publishes it itself only when that one runs out too or it has no such + successor. Before, the submitter of a transaction whose successor dropped it was the node that + published it. (#183) +- **A request body is charged against the admission budget at three times its length** (OPS-602; + node-local, no consensus or network change): `tests/burst_memory.rs` measures the 27 Oct burst + on a real node with a counting allocator and found a 5.8 MB body held about 18 MB while the + budget counted it once — 24 at once held 344 MiB on a 2 GB host's 117 MiB budget. The budget's + floor counts the largest body the same way (129 MiB on chain 21, which is now a 2 GB host's + default). Measured heap now stays at about the budget for any burst size; over-budget + submissions hear `-32005` before their body is read. New `deploy/probe-memory.sh` records each + host's RAM, swap, `MemoryHigh`, resident memory and budget. Not re-run on live 2 GB hosts; the + test's proofs do not verify, so verifier scratch is unmeasured. (#185) +- **From v0.9.0 a node gossips transactions on the anonymous lane only by default** (PRIV-601; + network only, no consensus change): the default follows the node's own version — a v0.8.x node + still subscribes and relays on the legacy signed topic unless `--anonymous-tx-only`, a v0.9.0 + node leaves it unless the new `--legacy-tx-relay` (or `--legacy-tx-publish`). Interoperates with + every node since v0.8.3. `tests/dandelion.rs` now runs a first-spy measurement in process; the + fleet run is still to be recorded, so origin hiding on chain 21 remains design intent. (#180) +- **`rand_getScanLeaves` is refused by default from v0.9.0** (RPC-201; RPC only, no consensus + change): the default now follows the node's own version — served before v0.9.0 unless + `--disable-scan-leaves`, refused from v0.9.0 unless the new `--enable-scan-leaves` — so the flip + the owner decided on 2026-11-05 happens with the version bump. No change on a v0.8.x node. A + serving node logs a warning at its 1st, 2nd, 4th, … served call (the count only), so the journal + shows whether wallets before v0.8.3 still scan through it. (#179) +- **`rand` refuses the retired RandProtocol prover keys** (VK-9; wallet CLI and pool router, no + consensus or network change): the build carries the list (the shared key retired 2026-10-01 and + the four members retired 2026-10-22), never stores a pairing with one and never sends one + anything, so a pairing file from an older `rand prover pair --trusted` is refused at the update + — run `rand prover pair --trusted` again. The pool answers a retired member's URL with HTTP 410 + and a reason instead of an empty 404. Wallets 0.6.9–0.8.0 still pin those keys and must be + updated; no release can change that. (#187) +- **A state proof that pays more than the exchange can hold is not pooled** (PERP-6, node policy + on every chain; consensus unchanged): the pool's pre-screen and its direct insert refuse a + `PerpStateProof` whose payouts total more than `Mempool::payout_ceiling` — the chain's count + under genesis `perps.bound_payouts`; otherwise `initial_collateral + rand_in − rand_out` for + RAND collateral and the token's total supply for a token — so it is not verified, pooled, + relayed or proposed by this node (`StateProofOverpays`: not cached, no peer penalised). On chain + 21 a block that carries one still applies; the validity rule is `bound_payouts`, at the chain-22 + cut. (#173) +- **`rand-node run --max-multisig-accounts `** (MSA-4, node policy; consensus unchanged): + once the multisig register holds that many accounts, the node's pool refuses a `CreateMultisig` + (`MultisigAtNodeCap`: not cached, no peer penalised), so it is not verified, pooled, relayed or + proposed there. Off by default. For a chain whose genesis sets no `multisig.max_accounts` (chain + 21): a block carrying a create still applies, so the register is bounded only when every + proposer sets the flag. The validity rule stays genesis `multisig.max_accounts`, an operator + input of the chain-22 cut. (#176) +- **Chain-22 cut: the record must say how validators hold the tier-18 key build** (CON-601, + CON-60; deploy script only): `deploy/cut-chain22-genesis.sh` refuses a cut record whose + `reason:`, `carries:` and `clients:` lines name neither `CON-601`, `validator memory` nor `key + build`. The engine pin bounds the shapes a state proof may declare, not what a key costs to + build: a block carrying a `max_tier`-18 proof makes every validator build a ~3.6 GB key, which + node policy cannot refuse. Listed as OPEN 9 in the script and in `docs/deploy.md`. (#174) +- **PERP-101, the `max_window_heights` decision** (docs and a test; no behaviour change): + `docs/perps.md` tabulates what the chain-22 input costs — heartbeat interval, state proofs a day + on an idle chain (`heights / max_window_heights`, so the lower bound 64 needs as many proofs as + chain 21's dense rule), back-pressure distance — and a ledger test pins the proof counts. (#177) +- **Prover pool router: an IPv6 /64 is one address, and active addresses split the capacity** + (VK-10, audit v13; `deploy/prover/router.py`, deploy only): the submit meter keys an IPv6 client + on its /64, and when more than 1/`SUBMIT_SHARE` addresses were admitted a submit in the last ten + minutes each one's refill is an equal split of the pool's capacity instead of `SUBMIT_SHARE` + each. Three addresses alone still take about 90% (README). Needs a router roll. (#178) + ## v0.8.6 — 2026-11-08 (chain 21, node-only; audit v13 rows; no wire, storage-format or consensus change; rolled 2026-11-09) Rolled 2026-11-09 00:16–02:22 UTC, one host at a time: tag `6591c884` (tagged 2026-11-08 19:20 UTC),