Fix JSON channel store flush race and missing rollback (#52)

[TYRMars]

Fixes #52.

Problem

JsonFileChannelBindingStore and JsonFileChannelInstanceStore mutated in-memory state under the write lock, then released the lock before flushing to disk. Two failure modes:

1. Durability race — two concurrent writers each dropped the lock before their atomic_write, so the disk writes could reorder (last write wins on disk). A change correct in memory could be durably lost, surfacing only after restart.
2. No rollback on flush error — the in-memory Vec was mutated first; a failing flush() propagated the error but left the mutation in place, so lookup/list reported a row that was never persisted.

Fix

Each write method now:

• builds the next state without touching the guard,
• flushes it while still holding the write lock (serialising disk writes so they cannot reorder), and
• commits the in-memory mutation only after the flush succeeds.

A failed flush leaves both memory and disk unchanged. Applied to upsert / delete / delete_for_conversation on the binding store and upsert / delete on the instance store.

Tests

Adds channel_binding_rolls_back_on_flush_failure, which forces a flush failure (occupies the atomic_write temp path with a directory) and asserts the in-memory state rolls back, then confirms a later write commits cleanly. cargo test -p harness-store and cargo clippy -p harness-store --all-targets -- -D warnings both pass.

https://claude.ai/code/session_01FhsDL2qEDL8MQnS7gmSdim

---

Generated by Claude Code