You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Integrate synchronized DKG PRF-key material and proof provenance #2459
Reuse compatible DKG implementation from the cryptographer's commit, adapting current crates/orchestration/keyshare, BFV and ZK owners rather than redesigning transport.
Carry directed pairwise PRF keys using the existing encrypted SK-share packing (residue + bit*q) under the synchronized DKG parameters. Specify canonical encoding, bit/residue constraints and unpacking.
Bind outgoing and incoming keys to their authenticated dealer/recipient identities and to the published committee's canonical IDs. Preserve private key material while exposing the commitments needed by synchronized C6.
Adapt setup circuits/witnesses/commitment links as required to establish sound provenance through DKG, including C1/C2/C3/C4 paths where applicable. Do not accept unanchored PRF keys at decryption time.
Persist/recover protocol-specific secret material with the E3 configuration. Keep the existing SS secret-shared smudging path functional.
Acceptance criteria
DKG produces matching directed PRF material for each required party pair, and restart restores the same authenticated material.
Tampered packing, party/key substitution, malformed fields and wrong-profile material are rejected at their owner boundaries.
All 15 PR reviews have been triaged, valid findings fixed, checked, committed/pushed and propagated
through the later slices without rewriting history. Existing reviews on #2492/#2494 were handled
without requesting another review. The other 13 full reviews ran sequentially under the included
one-review/hour allowance; no paid continuation was enabled.
This is review disposition, not CodeRabbit approval of every corrected head, human
cryptographic/protocol approval, merge authorization or activation. Corrected heads were not
automatically sent through another bot-review cycle.
Published stack and current checks
At 2026-11-18 03:39:12 UTC, all PRs are open and unmerged. GitHub check rollups on these exact
heads contain no failed, cancelled or pending checks. Skipped checks remain separate and are
not counted as passes. Adjacent bases, preserved ancestry and file counts were independently checked.
feat(aggregator)!: retain synchronized publication and all-H consent #2514: independently reproduced and fixed corrected pre-plan C1 loss with unchanged native bytes.
Added the missing final-consent flush-failure history after saved proof progress. The coverage
case already passed before the production fix; the C1 regression failed before it.
Each correction's PR carries its focused evidence and limits. Latest local correction checks include 113 aggregator tests, 178 keyshare tests, production-only compilation and focused Clippy.
At the final stack head, 5 publication histories and 13 synchronized keyshare tests passed.
Actor tests with synthetic proof bodies and mock flush adapters do not establish recursive-proof
validity or physical crash durability.
Current-head CI completed successfully for slice 14
and slice 15, including the selected
proof/recovery jobs. Slice 6's cancelled CI was traced to SDK proof-test ECONNRESET followed by the
workflow's cancellation step; its full exact-head rerun
succeeded. No source fix or diagnosis of the underlying network fault is claimed.
Bot tool timeouts/skips are not passes. The earlier linked-issue stack-completeness checks remain
inconclusive. feat(evm): recover and publish synchronized DKG intents #2515's summary mentions an unspecified, unverified restart edge without a concrete
finding or reproducer; no defect or universal restart assurance is inferred.
Synchronized admission remains disabled. Below-admission qualification is not admitted whole-node
activation, a reorg/rollback guarantee, transaction cancellation or distributed/across-restart
exactly-once publication.
Preserve the unreleased schema-9/protocol-9/generation-2 drain/reset cutover and gossip/sync 7/6.
Intermediate rolling upgrades and old-binary rollback over new journals remain unsupported.
SS/single-ciphertext behavior, complete fhe.rs v0.6.0 encodings, profile-owned selection, native
PRF ctx/eval, and deferred PRF tests are unchanged by these review corrections.
I independently checked the supplied review against the published heads, reproduced the valid
defects, and propagated ordinary fix commits/merges through the stack. No PR was merged, no profile
was enabled, and no deployment, release or artifact publication/restamping was performed.
Local VK-field/genealogy failures remain retryable generation errors; malformed witness inputs and independent expectation errors remain permanent. The effect-gate test now changes its payload so semantic dedup cannot mask a removed lifecycle guard.
The claimed SS metadata/committee leak is stale: the published 91fa73bd4 already filters SS authority, with constructor/live regressions. The reported-only early wrong-binding leaf wedge was reproduced across a cache-evicted Sled reopen and fixed: only provisional leaves are pruned when the first native plan arrives; fixed-plan validation stays strict.
Authenticated same-statement proof-byte repair is accepted before completion and retires stale work/correlations. Byte-identical redelivery does not reset work; completed local verdicts remain retained. Retained peer rows are candidates, not individually verified peer verdicts: Noir NodesFold recursively verifies all dealer proofs, and the executor verifies the resulting join. A failed join does not localize blame.
BFV-key removal again precedes setup cleanup, including on SS nodes, with accurate error context. A replayed differing/malformed candidate cannot block restoration of the validated exclusive saved setup. Loaded setups avoid redundant restore/write/flush cycles. Specific mutation checks now cover dealer-batch substitution and a coherent proof/setup PK substitution that reaches the independent native-PK check. The isolated reopen test now evicts Sled's cache.
Unrelated/superseded compute correlations are ignored. Local generation failures get one coalesced capped-backoff actor retry; permanent parameter errors do not. Recipient proof repair is coalesced rather than immediately redispatched for every new payload. Started owners still finish after demotion; canonical termination cancels retries. The shared compute gate permits actual fresh synchronized DKG/C5 attempts after a cached live failure, without changing SS outcome reuse. A production-only dependency check caught and fixed a dev-dependency leak. Admission-guard and storage-failure tests are separate; disk reopen evicts the cache.
Typed InvalidPublication() binding/forwarded provenance and terminal ThresholdNotMet() classification stop permanent retries. Live and recovered candidates use the same canonical validator; rejected candidates are logged/skipped before dedup, while journal read/decode/sequence errors remain fatal. Source tests now exercise the actual source owner independently of disabled admission; specific mutations include an otherwise correct payload signed by the wrong key.
Additional four findings
feat(dkg)!: authenticate synchronized recipient joins #2512 retry persistence — confirmed and fixed. Every recipient writer now validates at the
common flush boundary, not just initial persistence. A failed flush followed by authenticated
mixed genealogy genuinely failed after cache eviction/reopen before the fix; the valid partial
durable record now survives. This also preserves explicit below-admission qualification. CodeRabbit reply.
feat(keyshare)!: retain synchronized setup and publish recipient consent #2513 permanent restore failures — confirmed with qualification, fixed. Undecodable saved
bytes really scheduled automatic retries. A decodable wrong version already took a non-timer
validation-error path, but explicit drives repeated validation/reporting. Both now block automatic
setup work until operator repair/restart, preserving the saved evidence. Operational errors retain
capped, jittered backoff and one journal error per failure episode; later failures are traced.
Distinct injected error messages prove suppression is not merely journal deduplication. Invalid
new candidates remain replaceable. Identical-error dedup means a blanket claim of one actual
journal append per second was not independently established.
feat(aggregator)!: retain synchronized publication and all-H consent #2514 reverse-order C1 — confirmed and fixed. The corrected-first case
genuinely lost its slot before the fix. The existing validation owner now checks canonical
request/controller/config/committee binding, signer/slot and native PK commitment before retaining
C1, without waiting for the local plan. Both arrival orders preserve C5 readiness.
feat(aggregator)!: retain synchronized publication and all-H consent #2514 failed flush — confirmed and fixed. The final-consent regression
genuinely failed without redelivery before the fix. Validated progress now stays in memory with
effects fenced until acknowledgment, and a separate coalesced storage retry revalidates the latest
snapshot. Covering tests preserve live correlations, standby/demotion and canonical cancellation.
Third-round findings from the complete open-PR audit
The audit covered all 21 open PRs across seven repositories, including review bodies, hidden
nitpicks, outside-diff findings and resolved threads. CRISP #6 was draft/not reviewed; five other
non-Interfold PRs had no CodeRabbit review comments. Earlier resolved findings and the prior
15-PR disposition remain historical evidence, not clearance of later findings.
feat(dkg): add guarded synchronized proof workers #2510 (db9c8d381): child-proof shapes are checked against the loaded variant ABI;
malformed generated signals remain local generation failures while standalone verifiers return None; every child's independent semantics precede genealogy retry classification. The corrupt-VK
test now asserts the inner typed cause. Shape, generated-signal and both mixed-fault histories
failed before correction. Five inline findings received replies and were resolved.
feat(dkg)!: authenticate synchronized recipient joins #2512 (3aec1e89e): authenticated genealogy-only repair preserves every other statement
field; refused local/signed outputs have separate admission progress and automatic retry.
Missing recipient outcomes trigger fresh-correlated redelivery, with permanent-error suppression.
The genealogy wedge and missing-outcome histories failed before correction. Saturation coverage
leaves one admission slot on each attempt so durable dedup cannot hide duplicate local delivery.
The hidden roster nitpick is fixed: a valid alternative plan reaches the exact native-C4
immutability guard. Three remaining inline findings received replies and were resolved.
feat(aggregator)!: retain synchronized publication and all-H consent #2514 (630e1a372): stale errors cannot bypass running-job protection or answer a newer
attempt. A flushed replacement wrapper reaching the gate cancels superseded node/E3/correlation
attempts inside the shared pool; actual semaphore/Rayon queue tests preserve sibling work.
Executing native closures remain nonpreemptible and their late output is discarded.
Missing C5/wrapper outcomes redrive through the existing bounded retry. Bus refusal retains the
exact completed worker value instead of recomputing it. Known consent domains are checked before
replacement in live/replay intake; provisional peer anchors cannot prune early consent.
Pending consent still cannot authorize output before verified anchors. The stale-error, domain,
provisional-pruning and missing-outcome histories all failed before correction. The already-fixed
flush thread and inaccurate resolved README thread also received explicit dispositions.
The subsequent thread reconciliation found two new automatic-review findings on #2511
(C3 qualification, dealer admission retry).
Both reproduced before correction and are fixed in 9c6f5264a: C3 remains provisional until the
native plan and C2 can qualify its message commitment; mismatches release the slot while matching
proof bytes survive. Flushed dealer effects retain exact payload/correlation/causation across bus
refusal, with one pending retry per E3 and terminal cancellation. Downstream #2512 extends that
same pending queue to local and signed outputs; one-slot saturation proves the admitted local
output is not retried ahead of the remaining signed effect. The standalone #2511 prover library
passed 219 tests with two ignored, plus production checking.
Qualifications and remaining human gates
Schema compatibility still needs a human decision. Exactly 49 existing keyshare recovery
fixture rows change 4100→4121 bytes, including both SS profiles. Documentation now explicitly
separates unchanged SS peer payloads from changed SS recovery-row encodings. This remains the
disclosed unreleased schema9/protocol9/generation2 drain/reset cutover, final gossip/sync 7/6; sharing schema 9 does not support intermediate rolling upgrades or old-binary rollback.
No new migration, counter change, rollout or cryptographic approval is implied.
The recipient collector's PK/batch expectations are derived from authenticated input rows, not
independently from native PK bytes there. Its native C4/C2 provenance is independent; keyshare
subsequently checks retained native PK shares before consent. Early signed controller context
is provisional until native intent arrives; tests confirm incompatible early rows are pruned.
A faulty fixed-H member can withhold a proof or consent; no Byzantine fixed-H liveness guarantee
is claimed. Direct publication-actor retirement on CommitteeMemberExpelled remains an explicit
activation prerequisite, outside this disabled-admission follow-up. Registry active-H checks
still reject the publication; neither H nor the roster is silently reduced/replaced.
Once-per-boot republication of a completed result is intentional replay. PK chunks can be resent
after restart under the inherited SS helper; there is no distributed nonce/per-chunk exactly-once
journal. The public *_definition APIs are explicitly below-admission qualification APIs, not
production activation or an admission bypass.
The feat(dkg): add guarded synchronized proof workers #2510 outside-diff restart-after-KeyPublished concern remains stack-qualified: later
keyshare slices retain private setup and the confirmed_setup handoff, but admitted full-node
restart after publication remains an activation qualification gap. Terminal compute gating was
not weakened to reuse journaled outcomes. In-memory delivery recovery is not a transactional
outbox, distributed exactly-once guarantee or process-crash qualification.
The final combined library run at 078bf1678abff07c6e596b6024ef05930b0c665d passed 896 tests,
with 4 ignored (not counted as proof verification). The final head adds only a README
clarification and ordinary propagation merges; the checked production/test source is identical.
Package
Passed
Ignored
e3-zk-witness
22
0
e3-zk-prover
234
2
e3-keyshare
180
2
e3-aggregator
120
0
e3-multithread
38
0
e3-task-pool
5
0
e3-evm
293
0
e3-evm-contracts
4
0
Command: E3_ZK_PROVER_SKIP_FIXTURE_BUILD=1 cargo test --locked -p e3-zk-witness -p e3-zk-prover -p e3-keyshare -p e3-aggregator -p e3-multithread -p e3-task-pool -p e3-evm -p e3-evm-contracts --lib.
The same packages' production cargo check --locked ... --lib passes.
All six exact-head CI watchers completed successfully. Reconciliation at 2026-11-18 12:58:53 UTC
verified the listed remote heads, open/unmerged state, ancestry and file caps:
These are raw check counts, including successful duplicate checks. Concurrent base/head updates
cancelled duplicate runs; each cancelled check has completed equivalent checks on the same exact
head. Four cancelled job names per downstream PR were unexpanded matrix placeholders: their
expanded children passed, and the entire replacement ci workflow was independently verified
successful on that head. No failed, pending or unmatched cancelled checks were waived.
Skipped/neutral results and superseded cancellations are not successes or human review approval.
All inline threads on #2510–#2515 were resolved in the final thread snapshot.
Regression evidence includes genuine before-fix failures for local VK classification, early-leaf
recovery, dealer-proof replacement, SS key cleanup/saved-setup restoration, unrelated-result/retry
handling, permanent revert retirement and rejected recovered intents. Guard-removal mutations fail
for the effect gate, setup dealer batch, healthy-store admission guard, local source and wrong signer;
all production guards were restored.
Earlier pnpm evm:build and pnpm check:bindings passed; the ABI check retains the wrapper error's declaring
contract. Published SS artifacts were hydrated canonically with their original SOURCE_HASH=031f5d519d7d30bc; the source mismatch was disclosed, not restamped.
Ordinary focused Clippy (--lib --tests) passed for prover/witness/aggregator/multithread/task-pool
after the third round, and prover again after the final dealer corrections, with existing warnings
outside the correction files. Strict Clippy attempts are blocked by pre-existing needless_borrow / empty_line_after_doc_comments findings outside changed paths; they are not reported as green.
These actor/native/fake-process checks are not a new real-proof campaign, production admission,
secure-profile qualification, process-crash proof or successful deployment.
Historical network/indexer check failures remain undiagnosed despite successful reruns; no network
fault correction is claimed.
The runtime follow-up corrections change no durable/wire encoding, circuit statement, native
serialization, ABI, release counter or artifact. The compatibility/human-review limits above remain
unchanged, not resolved by these runtime fixes. A separate read-only diff/domain pass checked DKG,
proof-statement, persistence/replay and upgrade requirements against the post-change owners.
Open activation blocker: late outside-H local setup
The earlier restart disposition did not cover the live ordering. KeyPublished currently makes
the compute gate reject all synchronized DKG jobs, including a late Recipient join, and the
recipient collector treats the same stage as retirement of its unfinished local record. Keyshare
can reopen a setup that was already saved; that does not help an outside-H member whose join was
not dispatched or not yet complete. Admission is disabled and no live production qualification
is claimed.
Requested DKG follow-up: keep dealer/publication/consent retirement, and define a narrowly scoped
local continuation for an unfinished outside-H recipient against independently confirmed
synchronized authority and the exact fixed H batch. Do not make all N recipients a condition of
publication, change the roster, or globally reopen post-publication DKG. Account for any local
proof prerequisites, retained inputs and phase-specific cutoff that continuation requires.
Acceptance: in an effects-enabled history, deliver confirmed KeyPublished before an
outside-H recipient dispatch and, separately, while its join is in flight. In both cases show
durable local setup for that same batch and no consent or publication restart. Check terminal
failure/completion, wrong batch/owner, and reopen of pending versus already-saved setup separately.
This is an explicit #2461 prerequisite, not resolved by historical CI or a saved-setup reopen.
The lifecycle policy and evidence require protocol-owner review before activation.
The correction stack remains open and synchronized admission stays disabled. This updates the prior review evidence
for the newly reviewed cases; it is not protocol approval.
Retained the loaded Noir ABI checks. After canonical field decoding, local child shape mismatches (including this node's row in a mixed array) and inconsistent canonical fan-out are local generation failures; a malformed external row can still be a permanent input failure. The local C1 mismatch regression failed before the fix.
A C3 slot can hold one bounded alternate until native plan and C2 qualify it. wrong C3 → correct C3 → C2 finishes without redelivery, preserving the matching proof bytes. Qualification uses the same config/C2 statement for the backlog. Unchanged inputs only redrive unacknowledged writes; refused effects retain exact payload, correlation and causation behind one backed-off timer.
Different authenticated dealer bodies cannot supersede a running recipient join or a successful result. At most one alternate per active dealer slot; compatible repairs may activate after a matching reported failure, with a per-boot per-dealer cap. Genealogy repair retains the fixed statement. Local retries that haven't reported an outcome stay in flight; failure is not peer-invalid evidence. Storage and delivery retries are separate from the 60-second missing-outcome timer. New progress does not bypass storage backoff; refused bus delivery does not rewrite the store.
Invalid saved setup and a closed store now have different dispositions. An unexpectedly closed store requires checking storage health before restarting, not repair of saved bytes. Known shutdown starts no setup retry or repair report. Invalid saved state stays untouched; the operator guidance preserves the store/logs and matching key material. The late automatic review also caught refused consent without retry, SS completion blocked on synchronized storage, and removal failure hidden by a successful flush. All three reproduced and are fixed. The separate review-body backoff finding also reproduced: intervening setup drives no longer retry before the scheduled store retry.
KeyPublished still refuses new obsolete DKG work, but SS work already handed to the worker keeps its result and waiting correlations. Each live cancellation entry is owned by the forwarded mailbox item and worker future; cancelling a finished correlation leaves no marker. Executing native closures are still nonpreemptible. Publication validation reuses the native batch/PK commitment for all consent rows in the pass, without dropping context, signer, or exact-anchor checks.
Propagated through the writer without changing this round's writer code or its authority boundary. The PR description now discloses the intentional existing host-side ThresholdNotMet() terminal classification for SS publishCommittee; it is not an SS wire or contract change. Preflights/receipts still never install canonical authority.
Verification
cargo test --locked --no-fail-fast -p e3-zk-witness -p e3-zk-prover -p e3-keyshare -p e3-aggregator -p e3-multithread -p e3-task-pool -p e3-evm -p e3-evm-contracts -p e3-events -p e3-data --lib: 1,089 passed, four ignored on the final 4af0d20c6 tip, including 131 event and 54 storage tests.
The late feat(keyshare)!: retain synchronized setup and publish recipient consent #2513 tests count setup error attempts without event-bus deduplication, fence the event pipeline before inspecting consent history, await explicit flush-attempt acknowledgment at retry deadlines, and serialize the setup Sled cache evictions with the other synchronized disk histories.
Focused production checks and ordinary Clippy passed; existing Clippy warnings remain. Cargo runs used E3_ZK_PROVER_SKIP_FIXTURE_BUILD=1 to avoid unrelated fixture regeneration.
pnpm rust:test:sync-dkg:recovery with pinned E3_CUSTOM_BB: two passed. This included real recovered proof generation, wrapper acceptance and substitution rejection in Solidity, and the isolated registry writer/confirmed handoff scenario. This is not admitted full-node startup or a late outside-H history.
cargo check --locked -p e3-ciphernode-builder --lib and cargo test --locked -p e3-tests --test layout_lock passed. Normal push hooks passed, including both selected SS verifier comparisons and 1,724 relative documentation links. No generated artifacts changed.
Counts keep successful, skipped and neutral checks distinct. Superseded duplicate cancellations
have completed equivalent checks on the same SHA; no unpaired cancellation, failure or
pending check was treated as passing. All six adjacent-base ancestry checks pass and the file
counts include generated files (none changed in this correction round). No open inline thread
at the captured heads; prior replies were not repeated.
Actual-diff review followed DKG, proof statements, persistence and replay
and upgrades.
No new durable layout, peer encoding, circuit, verification key, deployed verifier or release
counter changes in this round. Remove is a local storage mailbox message, not a durable event.
SS single-ciphertext behavior, profile-owned selection, fhe.rs
v0.6.0 complete ciphertext encoding and native PRF ctx/eval remain intact; deferred PRF tests
are not counted as verification. The intentional SS writer retry change is noted above.
The unreleased schema 9 / protocol 9 / generation 2 drain/reset and final gossip/sync 7/6
remain. Earlier in the stack, 49 existing recovery layout-lock rows, including both SS profiles, change 4100 → 4121
bytes. Current-layout tests are not old-data replay/mixed-version compatibility. Intermediate
rolling upgrades and opening new state with the old binary remain unsupported; rollback needs the
prior compatible backup or controlled reset/resync. Human protocol/cryptographic and
rollout/rollback review is still required. The earlier SS archive keeps its original SOURCE_HASH=031f5d519d7d30bc; the source-hash mismatch remains disclosed, not restamped.
Activation remains blocked on late outside-H setup,
direct publication retirement on CommitteeMemberExpelled, and admitted full-node restart
qualification. Restoring a saved local setup does not cover one not dispatched or still in
flight at confirmed KeyPublished. The separate DKG lifecycle follow-up needs protocol-owner
agreement on exact confirmed H-batch authority, retained inputs and cutoff, without requiring
all N or reopening consent/publication. #2461
tracks the activation acceptance gate. The existing active-H contract check is retained.
Historical network/indexer failures were not diagnosed by this round and no network fix is claimed.
No C6/C7, decryption retry/failover, activation, or SAFE-refresh work was added. No PR was merged,
no new review was requested, and there was no deployment, release, or artifact publication.
Activity
refactor(bfv): limit synchronized foundation to DKG
chore(circuits): bring DKG-only foundation into leaf slice
chore(keyshare): integrate DKG-only foundation scope
neogiacomo commentedon Nov 18, 2026
CodeRabbit stack triage complete — admission remains disabled
All 15 PR reviews have been triaged, valid findings fixed, checked, committed/pushed and propagated
through the later slices without rewriting history. Existing reviews on #2492/#2494 were handled
without requesting another review. The other 13 full reviews ran sequentially under the included
one-review/hour allowance; no paid continuation was enabled.
This is review disposition, not CodeRabbit approval of every corrected head, human
cryptographic/protocol approval, merge authorization or activation. Corrected heads were not
automatically sent through another bot-review cycle.
Published stack and current checks
At 2026-11-18 03:39:12 UTC, all PRs are open and unmerged. GitHub check rollups on these exact
heads contain no failed, cancelled or pending checks. Skipped checks remain separate and are
not counted as passes. Adjacent bases, preserved ancestry and file counts were independently checked.
6b2c630c05a1f658e2d50b41bb0538011f7ff5580efeb703de8c4181b17eb2abeb756b0db74c48d2b5b547058a91fa73bd4699df051fc0ddc8d011e56e5f851513ddf76Maximum: 61 files, including generated outputs; every PR is below the 100-file cap.
Corrections and verification
calculation payload assertion.
including renamed relative paths and spaces.
until successful flush; replay/terminal and once-per-boot publication fences remain.
Added the missing final-consent flush-failure history after saved proof progress. The coverage
case already passed before the production fix; the C1 regression failed before it.
the earlier-owner BB correction found from test(keyshare): qualify synchronized DKG replay and publication #2506.
Each correction's PR carries its focused evidence and limits. Latest local correction checks include
113 aggregator tests, 178 keyshare tests, production-only compilation and focused Clippy.
At the final stack head, 5 publication histories and 13 synchronized keyshare tests passed.
Actor tests with synthetic proof bodies and mock flush adapters do not establish recursive-proof
validity or physical crash durability.
Current-head CI completed successfully for slice 14
and slice 15, including the selected
proof/recovery jobs. Slice 6's cancelled CI was traced to SDK proof-test
ECONNRESETfollowed by theworkflow's cancellation step; its full exact-head rerun
succeeded. No source fix or diagnosis of the underlying network fault is claimed.
Remaining gates and limits
this comment does not close Integrate synchronized DKG PRF-key material and proof provenance #2459.
inconclusive. feat(evm): recover and publish synchronized DKG intents #2515's summary mentions an unspecified, unverified restart edge without a concrete
finding or reproducer; no defect or universal restart assurance is inferred.
activation, a reorg/rollback guarantee, transaction cancellation or distributed/across-restart
exactly-once publication.
Intermediate rolling upgrades and old-binary rollback over new journals remain unsupported.
PRF
ctx/eval, and deferred PRF tests are unchanged by these review corrections.(Refresh fhe.rs SAFE dependency after the next Interfold release #2493), merging, deployment, release and artifact publication remain outside this work.
neogiacomo commentedon Nov 18, 2026
Third-round follow-up review: #2510–#2515
I independently checked the supplied review against the published heads, reproduced the valid
defects, and propagated ordinary fix commits/merges through the stack. No PR was merged, no profile
was enabled, and no deployment, release or artifact publication/restamping was performed.
db9c8d3814042222e9dbcbd2406ca969582193b49c6f5264a358749b33f9e991bd0df5f173be7af77471c5314db70e6bcfad0e44923a960a9a03aa88c00e4c2490b0b47e6931a837889757b279354b4ec3c14dd446d97fec5d00eed03c46cf7fd4967b4d4ab587ca3f43f4343edd44256187e58ca41a8891Disposition
91fa73bd4already filters SS authority, with constructor/live regressions. The reported-only early wrong-binding leaf wedge was reproduced across a cache-evicted Sled reopen and fixed: only provisional leaves are pruned when the first native plan arrives; fixed-plan validation stays strict.NodesFoldrecursively verifies all dealer proofs, and the executor verifies the resulting join. A failed join does not localize blame.InvalidPublication()binding/forwarded provenance and terminalThresholdNotMet()classification stop permanent retries. Live and recovered candidates use the same canonical validator; rejected candidates are logged/skipped before dedup, while journal read/decode/sequence errors remain fatal. Source tests now exercise the actual source owner independently of disabled admission; specific mutations include an otherwise correct payload signed by the wrong key.Additional four findings
common flush boundary, not just initial persistence. A failed flush followed by authenticated
mixed genealogy genuinely failed after cache eviction/reopen before the fix; the valid partial
durable record now survives. This also preserves explicit below-admission qualification.
CodeRabbit reply.
bytes really scheduled automatic retries. A decodable wrong version already took a non-timer
validation-error path, but explicit drives repeated validation/reporting. Both now block automatic
setup work until operator repair/restart, preserving the saved evidence. Operational errors retain
capped, jittered backoff and one journal error per failure episode; later failures are traced.
Distinct injected error messages prove suppression is not merely journal deduplication. Invalid
new candidates remain replaceable. Identical-error dedup means a blanket claim of one actual
journal append per second was not independently established.
genuinely lost its slot before the fix. The existing validation owner now checks canonical
request/controller/config/committee binding, signer/slot and native PK commitment before retaining
C1, without waiting for the local plan. Both arrival orders preserve C5 readiness.
genuinely failed without redelivery before the fix. Validated progress now stays in memory with
effects fenced until acknowledgment, and a separate coalesced storage retry revalidates the latest
snapshot. Covering tests preserve live correlations, standby/demotion and canonical cancellation.
Third-round findings from the complete open-PR audit
The audit covered all 21 open PRs across seven repositories, including review bodies, hidden
nitpicks, outside-diff findings and resolved threads. CRISP #6 was draft/not reviewed; five other
non-Interfold PRs had no CodeRabbit review comments. Earlier resolved findings and the prior
15-PR disposition remain historical evidence, not clearance of later findings.
db9c8d381): child-proof shapes are checked against the loaded variant ABI;malformed generated signals remain local generation failures while standalone verifiers return
None; every child's independent semantics precede genealogy retry classification. The corrupt-VKtest now asserts the inner typed cause. Shape, generated-signal and both mixed-fault histories
failed before correction. Five inline findings received replies and were resolved.
3aec1e89e): authenticated genealogy-only repair preserves every other statementfield; refused local/signed outputs have separate admission progress and automatic retry.
Missing recipient outcomes trigger fresh-correlated redelivery, with permanent-error suppression.
The genealogy wedge and missing-outcome histories failed before correction. Saturation coverage
leaves one admission slot on each attempt so durable dedup cannot hide duplicate local delivery.
The hidden roster nitpick is fixed: a valid alternative plan reaches the exact native-C4
immutability guard. Three remaining inline findings received replies and were resolved.
630e1a372): stale errors cannot bypass running-job protection or answer a newerattempt. A flushed replacement wrapper reaching the gate cancels superseded node/E3/correlation
attempts inside the shared pool; actual semaphore/Rayon queue tests preserve sibling work.
Executing native closures remain nonpreemptible and their late output is discarded.
Missing C5/wrapper outcomes redrive through the existing bounded retry. Bus refusal retains the
exact completed worker value instead of recomputing it. Known consent domains are checked before
replacement in live/replay intake; provisional peer anchors cannot prune early consent.
Pending consent still cannot authorize output before verified anchors. The stale-error, domain,
provisional-pruning and missing-outcome histories all failed before correction. The already-fixed
flush thread and inaccurate resolved README thread also received explicit dispositions.
concern was not reproduced and is not represented as a proven fix.
The subsequent thread reconciliation found two new automatic-review findings on #2511
(C3 qualification,
dealer admission retry).
Both reproduced before correction and are fixed in
9c6f5264a: C3 remains provisional until thenative plan and C2 can qualify its message commitment; mismatches release the slot while matching
proof bytes survive. Flushed dealer effects retain exact payload/correlation/causation across bus
refusal, with one pending retry per E3 and terminal cancellation. Downstream #2512 extends that
same pending queue to local and signed outputs; one-slot saturation proves the admitted local
output is not retried ahead of the remaining signed effect. The standalone #2511 prover library
passed 219 tests with two ignored, plus production checking.
Qualifications and remaining human gates
fixture rows change 4100→4121 bytes, including both SS profiles. Documentation now explicitly
separates unchanged SS peer payloads from changed SS recovery-row encodings. This remains the
disclosed unreleased schema9/protocol9/generation2 drain/reset cutover, final gossip/sync
7/6; sharing schema 9 does not support intermediate rolling upgrades or old-binary rollback.
No new migration, counter change, rollout or cryptographic approval is implied.
independently from native PK bytes there. Its native C4/C2 provenance is independent; keyshare
subsequently checks retained native PK shares before consent. Early signed controller context
is provisional until native intent arrives; tests confirm incompatible early rows are pruned.
is claimed. Direct publication-actor retirement on
CommitteeMemberExpelledremains an explicitactivation prerequisite, outside this disabled-admission follow-up. Registry active-H checks
still reject the publication; neither H nor the roster is silently reduced/replaced.
after restart under the inherited SS helper; there is no distributed nonce/per-chunk exactly-once
journal. The public
*_definitionAPIs are explicitly below-admission qualification APIs, notproduction activation or an admission bypass.
keyshare slices retain private setup and the
confirmed_setuphandoff, but admitted full-noderestart after publication remains an activation qualification gap. Terminal compute gating was
not weakened to reuse journaled outcomes. In-memory delivery recovery is not a transactional
outbox, distributed exactly-once guarantee or process-crash qualification.
review evidence exists in the earlier consolidated comment. That is not human approval, and
these new corrections still require review. No new bot review was requested, especially not on
feat(bfv): add synchronized DKG key-transport foundation #2492/feat(circuits): add isolated synchronized DKG leaves #2494. Main was not incidentally merged/rebased.
Verification
The final combined library run at
078bf1678abff07c6e596b6024ef05930b0c665dpassed 896 tests,with 4 ignored (not counted as proof verification). The final head adds only a README
clarification and ordinary propagation merges; the checked production/test source is identical.
e3-zk-witnesse3-zk-provere3-keysharee3-aggregatore3-multithreade3-task-poole3-evme3-evm-contractsCommand:
E3_ZK_PROVER_SKIP_FIXTURE_BUILD=1 cargo test --locked -p e3-zk-witness -p e3-zk-prover -p e3-keyshare -p e3-aggregator -p e3-multithread -p e3-task-pool -p e3-evm -p e3-evm-contracts --lib.The same packages' production
cargo check --locked ... --libpasses.All six exact-head CI watchers completed successfully. Reconciliation at 2026-11-18 12:58:53 UTC
verified the listed remote heads, open/unmerged state, ancestry and file caps:
These are raw check counts, including successful duplicate checks. Concurrent base/head updates
cancelled duplicate runs; each cancelled check has completed equivalent checks on the same exact
head. Four cancelled job names per downstream PR were unexpanded matrix placeholders: their
expanded children passed, and the entire replacement
ciworkflow was independently verifiedsuccessful on that head. No failed, pending or unmatched cancelled checks were waived.
Skipped/neutral results and superseded cancellations are not successes or human review approval.
All inline threads on #2510–#2515 were resolved in the final thread snapshot.
Regression evidence includes genuine before-fix failures for local VK classification, early-leaf
recovery, dealer-proof replacement, SS key cleanup/saved-setup restoration, unrelated-result/retry
handling, permanent revert retirement and rejected recovered intents. Guard-removal mutations fail
for the effect gate, setup dealer batch, healthy-store admission guard, local source and wrong signer;
all production guards were restored.
Earlier
pnpm evm:buildandpnpm check:bindingspassed; the ABI check retains the wrapper error's declaringcontract. Published SS artifacts were hydrated canonically with their original
SOURCE_HASH=031f5d519d7d30bc; the source mismatch was disclosed, not restamped.Ordinary focused Clippy (
--lib --tests) passed for prover/witness/aggregator/multithread/task-poolafter the third round, and prover again after the final dealer corrections, with existing warnings
outside the correction files. Strict Clippy attempts are blocked by pre-existing
needless_borrow/empty_line_after_doc_commentsfindings outside changed paths; they are not reported as green.These actor/native/fake-process checks are not a new real-proof campaign, production admission,
secure-profile qualification, process-crash proof or successful deployment.
Historical network/indexer check failures remain undiagnosed despite successful reruns; no network
fault correction is claimed.
The runtime follow-up corrections change no durable/wire encoding, circuit statement, native
serialization, ABI, release counter or artifact. The compatibility/human-review limits above remain
unchanged, not resolved by these runtime fixes. A separate read-only diff/domain pass checked DKG,
proof-statement, persistence/replay and upgrade requirements against the post-change owners.
Domain review references: DKG,
proof statements,
persistence/replay,
upgrades.
Required human protocol/compatibility review remains outstanding.
neogiacomo commentedon Nov 18, 2026
Open activation blocker: late outside-H local setup
The earlier restart disposition did not cover the live ordering.
KeyPublishedcurrently makesthe compute gate reject all synchronized DKG jobs, including a late
Recipientjoin, and therecipient collector treats the same stage as retirement of its unfinished local record. Keyshare
can reopen a setup that was already saved; that does not help an outside-H member whose join was
not dispatched or not yet complete. Admission is disabled and no live production qualification
is claimed.
Requested DKG follow-up: keep dealer/publication/consent retirement, and define a narrowly scoped
local continuation for an unfinished outside-H recipient against independently confirmed
synchronized authority and the exact fixed H batch. Do not make all N recipients a condition of
publication, change the roster, or globally reopen post-publication DKG. Account for any local
proof prerequisites, retained inputs and phase-specific cutoff that continuation requires.
Acceptance: in an effects-enabled history, deliver confirmed
KeyPublishedbefore anoutside-H recipient dispatch and, separately, while its join is in flight. In both cases show
durable local setup for that same batch and no consent or publication restart. Check terminal
failure/completion, wrong batch/owner, and reopen of pending versus already-saved setup separately.
This is an explicit #2461 prerequisite, not resolved by historical CI or a saved-setup reopen.
The lifecycle policy and evidence require protocol-owner review before activation.
neogiacomo commentedon Nov 18, 2026
Fifth-round corrections: #2510–#2515
The correction stack remains open and synchronized admission stays disabled. This updates the prior
review evidence
for the newly reviewed cases; it is not protocol approval.
wrong C3 → correct C3 → C2finishes without redelivery, preserving the matching proof bytes. Qualification uses the same config/C2 statement for the backlog. Unchanged inputs only redrive unacknowledged writes; refused effects retain exact payload, correlation and causation behind one backed-off timer.KeyPublishedstill refuses new obsolete DKG work, but SS work already handed to the worker keeps its result and waiting correlations. Each live cancellation entry is owned by the forwarded mailbox item and worker future; cancelling a finished correlation leaves no marker. Executing native closures are still nonpreemptible. Publication validation reuses the native batch/PK commitment for all consent rows in the pass, without dropping context, signer, or exact-anchor checks.ThresholdNotMet()terminal classification for SSpublishCommittee; it is not an SS wire or contract change. Preflights/receipts still never install canonical authority.Verification
cargo test --locked --no-fail-fast -p e3-zk-witness -p e3-zk-prover -p e3-keyshare -p e3-aggregator -p e3-multithread -p e3-task-pool -p e3-evm -p e3-evm-contracts -p e3-events -p e3-data --lib: 1,089 passed, four ignored on the final4af0d20c6tip, including 131 event and 54 storage tests.E3_ZK_PROVER_SKIP_FIXTURE_BUILD=1to avoid unrelated fixture regeneration.pnpm rust:test:sync-dkg:recoverywith pinnedE3_CUSTOM_BB: two passed. This included real recovered proof generation, wrapper acceptance and substitution rejection in Solidity, and the isolated registry writer/confirmed handoff scenario. This is not admitted full-node startup or a late outside-H history.cargo check --locked -p e3-ciphernode-builder --libandcargo test --locked -p e3-tests --test layout_lockpassed. Normal push hooks passed, including both selected SS verifier comparisons and 1,724 relative documentation links. No generated artifacts changed.Exact-head GitHub checks
7dbfa7e573470e75fffee546306d1b20e5f9964c8dc574af0d20c6Counts keep successful, skipped and neutral checks distinct. Superseded duplicate cancellations
have completed equivalent checks on the same SHA; no unpaired cancellation, failure or
pending check was treated as passing. All six adjacent-base ancestry checks pass and the file
counts include generated files (none changed in this correction round). No open inline thread
at the captured heads; prior replies were not repeated.
On #2513's final head, the
synchronized DKG proof job
and the recovery and registry qualification job
both completed successfully. No result here enables production admission.
Compatibility and activation limits
Actual-diff review followed DKG,
proof statements,
persistence and replay
and upgrades.
No new durable layout, peer encoding, circuit, verification key, deployed verifier or release
counter changes in this round.
Removeis a local storage mailbox message, not a durable event.SS single-ciphertext behavior, profile-owned selection, fhe.rs
v0.6.0 complete ciphertext encoding and native PRF
ctx/evalremain intact; deferred PRF testsare not counted as verification. The intentional SS writer retry change is noted above.
The unreleased schema 9 / protocol 9 / generation 2 drain/reset and final gossip/sync 7/6
remain. Earlier in the stack, 49 existing recovery layout-lock rows, including both SS profiles, change 4100 → 4121
bytes. Current-layout tests are not old-data replay/mixed-version compatibility. Intermediate
rolling upgrades and opening new state with the old binary remain unsupported; rollback needs the
prior compatible backup or controlled reset/resync. Human protocol/cryptographic and
rollout/rollback review is still required. The earlier SS archive keeps its original
SOURCE_HASH=031f5d519d7d30bc; the source-hash mismatch remains disclosed, not restamped.Activation remains blocked on late outside-H setup,
direct publication retirement on
CommitteeMemberExpelled, and admitted full-node restartqualification. Restoring a saved local setup does not cover one not dispatched or still in
flight at confirmed
KeyPublished. The separate DKG lifecycle follow-up needs protocol-owneragreement on exact confirmed H-batch authority, retained inputs and cutoff, without requiring
all N or reopening consent/publication. #2461
tracks the activation acceptance gate. The existing active-H contract check is retained.
Historical network/indexer failures were not diagnosed by this round and no network fix is claimed.
No C6/C7, decryption retry/failover, activation, or SAFE-refresh work was added. No PR was merged,
no new review was requested, and there was no deployment, release, or artifact publication.