Explore write-path hardening, batching, and agent coordination

[justrach]

Context

Recent agent-driven experiments have been useful for exercising TurboDB as both a document store and a coordination substrate. There are a few related areas worth exploring together because they all touch the write path, recovery behavior, and multi-agent workloads.

Areas to explore

• Make the WAL recovery path more explicit in normal startup: verify committed document entries are replayed into the right collection, and add tests that kill/restart around pending writes.
• Add crash-atomic markers for collection transactions so an interrupted batch cannot replay as a partial batch after restart.
• Improve the bulk write path so it behaves like a real batch operation instead of a best-effort loop over single inserts.
• Decide whether the existing replication/sequencer modules should become part of the active server write path, or remain experimental until the integration is complete.
• Keep developing the agent coordination harness as a workload for retention, atomic claim/result writes, shared chat/file coordination, and tool-call auditability.

Suggested first pass

• Add a focused recovery test that starts TurboDB, writes via HTTP, stops/restarts the process, and verifies values are retained.
• Add a transaction crash/replay test around /db/:col/txn.
• Keep the agent experiment harness in a branch so we can run and compare coordination modes without mixing it into the core engine path yet.

Notes

This is intentionally exploratory. The goal is to identify which changes belong in the engine, which belong in tests, and which should stay as separate agent workload tooling.