fix(vfs): stop wedging the level-1 tailer on a straddling first compaction

[yzxcj797]

Fixes #1460.

Investigation (reproduced from the report's timeline)

rebuildIndex seeds the level-1 tailer position from an L0-derived value when the index is built while L1 is still empty:

maxTXID1 := maxLevelTXID(infos, 1)
if maxTXID1 == 0 {
    maxTXID1 = pos.TXID   // e.g. 6 — NOT an L1 file boundary
}

The writer's first L0→L1 compaction then emits a file whose range straddles that value (1..0xb around position 6). The level-1 seek filters by MinTXID, so the straddling file is invisible forever; when the next L1 file (0xc..0x1d) appears, the strict contiguity check fails on every poll and the follower freezes at TXID 6 while storage remains fully contiguous:

poll L1: non-contiguous ltx file: level=1, current=0000000000000006, next=000000000000000c-000000000000001d

Fix

maxTXID1 stays zero when there are no L1 files, so the first L1 file — which starts at TXID 1 regardless of how far L0 has advanced — is consumed contiguously, and the position tracks real L1 boundaries from then on.

Scope

• This PR changes only the seeding in rebuildIndex (plus tests). It does not change the MinTXID-based seek semantics of any storage backend, nor the pollLevel contiguity rules — with a correctly-zero seed neither can encounter a straddle.
• The vfs-tagged test suite did not compile on main (the in-file test mocks predate the SetLogger addition to ReplicaClient); the three mocks get a no-op SetLogger so the tagged suite builds again. Pre-existing failures unrelated to this change (e.g. TestReplica_SyncOnceLimitsLTXFiles, TestDB_SyncChunksWALAtCommitBoundary) reproduce identically with and without this fix.

Tests

Updated TestVFSFile_OpenSeedsLevel1PositionFromPos (which pinned the old seeding) to TestVFSFile_OpenKeepsLevel1PositionAtZeroWithoutL1Files, and added TestVFSFile_PollConsumesStraddlingL1Compaction, which replays the report's exact timeline: open with L0 at 6 and L1 empty, then present L1 1..0xb and 0xc..0x1d, then poll.

go test -tags vfs -run 'TestVFSFile_PollConsumesStraddlingL1Compaction|TestVFSFile_OpenKeepsLevel1PositionAtZero' -count=1 .
# ok

Differential: reverting only the seeding (tests kept) fails with exactly the report's wedge error — poll L1: non-contiguous ltx file: level=1, current=0000000000000006, next=000000000000000c-000000000000001d.

[darkgnotic]

Thank you @yzxcj797 . 🙏

FWIW, I asked Claude about this solution and it seemed to think that it would break compaction in write mode (e.g. introduced in #953). But TBH, I would love for this to be the correct solution, as it's much simpler. 🤞

[yzxcj797]

@darkgnotic thanks for flagging that — I checked the write-mode path against #953 and I don't see an interaction:

• maxTXID1 is consumed in exactly one place: poll() → pollLevel(ctx, 1, f.maxTXID1, ...) (vfs.go:2511/2538), the read-side L1 tailer. The only other reference is the MaxTXID1() accessor, which has no callers outside tests.
• The write-mode compactor (compactor.go) never reads f.maxTXID1; its level/retention decisions come from the replica client listing and its own state, so a zero seed can't steer compaction.
• In write mode the poll goroutine still runs ("in case another writer creates the database"). With the zero seed it consumes the first L1 compaction output (one-time, bounded by the first compaction's files) instead of skipping past it; from then on it tracks real L1 boundaries exactly as before. Page-index fetches that hit a file being compacted away fail the poll and retry next tick — the same failure/retry mode the pre-existing fresh-consumption path already has.

So the effect is strictly within the read-side tailer that the issue is about.