The indexes file is intentionally not fsynced, so corruption after a
crash is expected. Previously this caused a badmatch crash in
find_snapshots during init. Now ra_snapshot carries the machine config
and can recover live indexes by reading the snapshot and calling
ra_machine:live_indexes/2, the same approach used in complete_accept.

When a crash occurs after the segment writer flushes a WAL file to
segments but before the WAL file is deleted, recovery replays the
same WAL creating segments that overlap with those from the first
flush. compact_segrefs correctly handles this by truncating the
range of partially overlapping segment refs. However the deletion
logic used the -- operator which compares full {Filename, Range}
tuples. A segment whose range was truncated (but not removed) no
longer matched its original ref, so it appeared in the diff and
was deleted even though the reader still referenced it. The
subsequent fold during state machine recovery then crashed with
ra_log_failed_to_open_segment enoent.

Compare by filename only when deciding which segments to delete,
so that segments still referenced by the reader (even with a
truncated range) are preserved.

When a leader receives a pre_vote_rpc from a follower with a stale
term, make_all_rpcs now includes peers in snapshot_backoff status
alongside normal peers. This ensures the lagging follower that
triggered the pre-vote gets its snapshot expeditiously rather than
waiting for the backoff timer to fire. The pending backoff timer is
cancelled via a new cancel_snapshot_retry_timer effect before the
RPC is sent.

…concurrently

During multi-file WAL recovery after a power-off, the segment writer
processes mem tables from earlier WAL files asynchronously. When servers
have no Pid (normal during recovery), the segment writer deletes entries
directly from the mem table ETS. If this deletion races with recovery of
the next WAL file, recover_entry calls mem_table_please which re-scans
the (now partially depleted) ETS table. The resulting ra_mt state has a
LastSeq that no longer matches the PrevIdx tracked in the writers map,
causing ra_mt:insert_sparse to return {error, gap_detected} — an
unhandled case_clause in recover_entry that crashes the node at boot.

Fix by carrying the Tables map across WAL files in the recovery fold,
alongside the already-carried Writers map. This way recover_entry reuses
the ra_mt state it built during earlier file recovery rather than
re-scanning a potentially mutated ETS table.

Made-with: Cursor

Else it may fail to boot. Ignore for windows.

New servers should register _after_ log initialisation to ensure the
config file is fully written as it is required for successful recovery

Improve crash recovery resilience