[citizen-stig]

Every existing fuzz target drives the database from one thread with one session at a time. Nothing exercises coexisting sessions, commits waiting on them, or rollbacks racing with sessions on other threads.

This adds two targets with a shared harness (fuzz_targets/concurrent/mod.rs). Each round releases 2–4 threads. Each thread runs sessions with reads, writes, proofs, nested sessions and Nomt::read, and ends each session with a blocking, non-blocking or overlay commit, or a drop. Pauses widen the race windows. Commit concurrency, I/O workers and warm-up are randomized, and a round may reopen the database.

• concurrent_sessions: commits race each other for the write lock.
• concurrent_rollback: adds rollbacks from any thread with a 1–8 entry rollback log. Commits and rollbacks queue on a harness lock (never taken while holding a session), so each rollback's outcome is known.

Checks:

• a session's reads, proofs, nested sessions and Nomt::read under it match the state its prev root names
• a finished session's root matches a root computed from the model
• a commit fails only as stale, and only once the root has moved
• successful commits and rollbacks chain from the empty trie to the current root
• the database matches the model after every round and after every reopen

These targets found the bugs fixed in #973, #974 and #975, so they fail until those land.

🤖 Generated with Claude Code