fix(archiver): serve log-by-tag queries from read-only snapshots

[spalladino]

Stack

Part of A-1817. Bottom to top, each PR targets the branch below it:

• fix(archiver): avoid loading full blocks during checkpoint ingestion #25279 spl/faster-block-ingestion -> merge-train/spartan-v5 — cheaper checkpoint ingestion, so writes hold the store for less time
• feat(kv-store): expose read-only lmdb transactions #25280 spl/kv-store-read-only-tx -> spl/faster-block-ingestion — adds readOnlyTransaction plus txId on the native GET / START_CURSOR
• fix(archiver): batch multi-key point reads into one LMDB round trip #25282 spl/batched-kv-gets -> spl/kv-store-read-only-tx — adds getMany / getManyAsync; depends on feat(kv-store): expose read-only lmdb transactions #25280 for txId on GetRequest
• fix(archiver): speed up log-by-tag queries #25254 spl/faster-get-private-logs-by-tags -> spl/batched-kv-gets — faster tag scans; depends on fix(archiver): batch multi-key point reads into one LMDB round trip #25282 for getManyAsync on the log store's per-block reads
• fix(archiver): serve log-by-tag queries from read-only snapshots #25315 spl/log-store-read-only-snapshots -> spl/faster-get-private-logs-by-tags — moves the log store's reads onto snapshots; depends on feat(kv-store): expose read-only lmdb transactions #25280 for readOnlyTransaction <- this PR

Context

Fixes A-1817.

#25254 kept tag log queries inside db.transactionAsync to preserve snapshot consistency, which on lmdb-v2 means routing them through the store's single serial writer queue: a query waits for every queued write (notably checkpoint ingestion) to commit before it even starts. #25280 added readOnlyTransaction, which gives the same one-snapshot guarantee without touching the writer queue, so the log store can have consistency and not queue.

Approach

• LogStore's read paths (getPrivateLogsByTags, getPublicLogsByTags, getPrivateLogsForBlock, getPublicLogsForBlock) switch from db.transactionAsync to db.readOnlyTransaction. The inner reads already resolve the ambient transaction through acquireReadTx, so these are one-line swaps. addLogs and deleteLogs are untouched.

• The class JSDoc no longer justifies staying on the writer queue; it documents the snapshot instead.

• TAG_SCAN_CONCURRENCY drops from 8 to 2. Cursors bound to one read-only snapshot serialize on that snapshot's mutex in C++, so a wide pool only hides the per-cursor round trip. Measured on a seeded store (20 blocks x 4 txs x 5 private logs, 100-tag query, 30 iterations after warmup, two runs):

  TAG_SCAN_CONCURRENCY	avg latency	avg with 5 queries in flight
  1	7.51 / 7.32 ms	26.3 / 26.9 ms
  2	4.72 / 5.10 ms	26.9 / 27.8 ms
  4	4.89 / 5.00 ms	27.9 / 27.8 ms
  8	4.93 / 5.08 ms	27.9 / 28.2 ms
  16	5.15 ms	29.3 ms

  Going from 1 to 2 buys a third of the latency; everything above 2 is flat uncontended and marginally worse under concurrency, while holding more cursors.

• Each open snapshot holds one of the LMDB reader slots, so MAX_READERS is no longer hardcoded in kv-store/src/lmdb-v2/factory.ts: it comes from DataStoreConfig.dataStoreMaxReaders (DATA_STORE_MAX_READERS), defaulting to the previous 16.

Testing

log_store_write_contention now asserts the read does not wait for the queued writes. With 5 x 50ms writes queued in front of it:

• on transactionAsync (before): baseline 10.56ms, contended 259.92ms — fails
• on readOnlyTransaction (after): baseline 11.40ms, contended 8.39ms — passes

The bound is contendedMs < baselineMs + queuedWriteMs / 2, leaving 125ms of jitter headroom so it only trips when the read is genuinely serialized.