fix(defaults): async state-commit for full-mode followers (PLT-537)

[bdchatham]

Problem

Full-mode followers (node/rpc/syncer) default to synchronous memIAVL commit (sc-async-commit-buffer=0), which holds the consensus lock cs.mtx across the entire block write. That starves the single-goroutine consensus StateChannel drain → dropped NewRoundStep/HasVote gossip → degraded peer-state view → bistable fall-behind (recurring NodeFellBehind on pacific-1). Root cause: PLT-537.

Change

Default full mode to AsyncCommitBuffer = 100 in applyFullOverrides (matches the already-async archive mode). Flows to full + archive (archive inherits applyFullOverrides); validators and seeds stay synchronous by construction (separate override paths) — async commit's in-memory crash window is unacceptable for a signing node.

Adds TestDefaultForMode_AsyncCommitBufferByMode pinning full=100, archive=100, validator=0, seed=0 so the fix and the validator-safety constraint can't silently regress.

Validation

• Canary on pacific-1 node-0: 23.7k blocks behind → tip, 0 StateChannel drops, block_processing 45→29 ms, held 18 min.
• Full pacific-1 K8s follower fleet (5 nodes) recovered to tip; unchanged sync-commit nodes stayed trapped (syncer-0-0 relapsed → restart-alone is not durable, the config is).
• EC2 state-sync-node-0: 32.5k → tip, block_processing 194→11 ms, drops → 0 (cross-topology confirmation).

Rollout

This is a library default — it reaches prod once consumed downstream: sei-config tag → sei-node-controller go.mod bump → seictl sidecar rebuild → release → rollout (pod recreation re-renders app.toml). Manual sc-async-commit-buffer=100 patches are holding the live fleet in the meantime.

Risk

Async commit buffers un-flushed commits in memory; on an ungraceful crash the tail is re-fetched via WAL/blocksync (no corruption). Appropriate for followers/RPC/archive; explicitly not validators (preserved here).

Refs: PLT-537

🤖 Generated with Claude Code

[cursor]

PR Summary

Medium Risk
Changes default state-commit behavior for all new full/archive deployments; async commit trades a small in-memory crash window for lower consensus lock hold time, which is intentional for non-signing nodes only.

Overview
Full-mode defaults now set Storage.StateCommit.AsyncCommitBuffer = 100 in applyFullOverrides, so follower/RPC-style nodes use async memIAVL commits.id commit instead of synchronous (0). Archive still ends at 100 via applyFullOverrides plus its existing override; validator and seed remain 0 (no change on their override paths).

Adds TestDefaultForMode_AsyncCommitBufferByMode to lock full=100, archive=100, validator=0, seed=0 so the follower default and signing-node constraint do not regress silently.

^{Reviewed by Cursor Bugbot for commit 7e5d431. Bugbot is set up for automated code reviews on this repo. Configure here.}