fix(service): preserve invalidated slots in the initial-round DKG node list

[0xHansLee]

Problem

The initial-round kyber node list was built from VERIFIED-only registrations. extractSortedPubKeys assigns each kyber position as a dense index sorted by reg.Index, so dropping a member invalidated mid-dealing shifts every later node's index. A node that first builds its DistKeyGenerator after an invalidation gets a shorter, shifted list that diverges from peers who built earlier — and fails the len(regs) == network.Total check, since network.Total is fixed at BeginDealing and never decremented.

Fix

Build (and validate) the node list from all registrations (any status) via GetAllRegisteredDKGRegistrations, mirroring the resharing path. Invalidated members keep their slot, so the list has a stable length equal to network.Total regardless of when a node builds. PID matching (matchPIDFromRegistrations/CachePID) is status-insensitive — it keys off each registration's own DkgPubKey and on-chain Index — so the retained slots are harmless.

Tests

• TestFetchRoundContext_PreservesInvalidatedSlot

issue: none

[jinn-agent]

The core fix is sound: swapping GetAllParticipantDKGRegistrations (VERIFIED-only) for GetAllRegisteredDKGRegistrations (all statuses) in fetchRoundContext mirrors the already-working resharing path and keeps the kyber position list stable regardless of when a node builds after a mid-dealing invalidation. The accompanying test covers the regression scenario and the negative (shrunk-list) case correctly. One assumption implicit in the fix is worth calling out, and the stub fallback has a subtle edge case noted below.

---

Review iteration 2 · Commit 527faac · 2026-07-22T08:56:59Z

[yingyangxu2026]

The commit message says "Registrations stays VERIFIED-only for status-sensitive consumers," but in the diff rc.Registrations now carries all statuses — the same registrations variable feeds both SortedPubKeys and rc.Registrations, and the test asserts Len(Registrations, 3) after an invalidation. The message is stale / contradicts the code; worth updating so it doesn't mislead reviewers.