I Shipped...

I made generated workspace YAML quote and escape dependency paths safely, including Windows paths with backslashes.

I kept selected fields, aliases and source details intact as relation queries are combined, ordered and updated.

I made the test entry point build and seal Inspector assets before either test lane starts, preventing missing files and rebuilds from racing with browser tests.

I made correctness snapshot tests handle the file-collision behaviour seen on macOS and made nested fixture checks work on other platforms.

I kept messages in send order on each simulated network link even when random delays would otherwise let a later message arrive first.

I made a receiver finish its initial setup from payload references it already has locally, without waiting for another update to arrive.

I made manual test runs fail clearly when a requested starter does not exist, instead of silently skipping every test and reporting success.

I added regression checks showing that nested records are admitted correctly and that their allocation use grows within a defined bound as nesting gets deeper.

I made exclusive reads compare the rows their filter selected, so a change to an unrelated row in the same table no longer causes a false conflict.

I wired the Inspector into SvelteKit development and made the managed runtime shut down cleanly when the plugin closes or rebuilds.

I made external read readiness wake the owning server or relay before it ticks subscribers, so waiting work resumes without unsafe repeated scheduling.

I made the React and Next starters show local delete failures and finish the deleting indicator after local saving, while remote syncing continues separately.

I gave compressed messages separate size limits before and after decoding, so legal compressed data is accepted while oversized declarations are rejected early.

I added bounded reuse for complete row history and eligible query plans, without sharing query results or access rights between requests.

I kept the schema identity that originally described a read, so unrelated schema changes or a renamed table cannot make the read check the wrong table.

I made raw writes reject records stored under a key that disagrees with the record's own primary key, before the inconsistent data can be saved.

I made permanent transport send errors reach the caller and kept quick reconnect confirmations from being lost while older messages are drained.

I made cancellation release a read that never finishes, allowing the next queued operation to proceed instead of waiting behind it forever.

I made explicit relationship declarations reject nested arrays as reference targets early, while still allowing nested arrays as ordinary data.

I made a restarted browser recover pending offline edits even when their earlier history is temporarily incomplete, and replay long histories without overflowing the stack.