Fireside 0.1.0-next.6

Write-behind durability is the default (#39). A Firestore commit is acknowledged after its journal write and a non-durable redb commit; a background flusher (every second) and shutdown sync the journal, persist the commits with one durable redb commit and truncate the journal. Storage keeps the same contract with its own metadata journal and syncs uploaded files before the metadata that references them becomes durable. Killing the process loses nothing acknowledged (reopen replays the journal); only a kernel crash or power cut can lose the last interval. --durability per-commit restores the previous flush-per-write behaviour. Measured on a Mac: single document set 13.1 ms → 0.59 ms. In the consumer's Linux soak the write-commit median went from 23.9 ms with per-commit durability to 3.2 ms (official emulator on the same host and workload: 197.5 ms) and the 64 KiB Storage upload from 18.0 ms to 2.7 ms.

Every HTTP front (REST, Storage, Auth, hub, UI, Requests feed) sets TCP_NODELAY on accepted connections, as the gRPC front already did (#35). A 64 KiB download over a reused keep-alive connection on Linux fell from 40.7 ms to 0.5 ms (Nagle waiting on the client's delayed ACK); fresh connections were unaffected.

Queries not served in __name__ order no longer decode every scanned document (#38). Stored values gain a field directory that lets filters, order keys, cursors and vector distances read one field path; only result pages are decoded. __name__ ==/in filters read the named documents directly and count-only aggregations use a dedicated count path. Values written by earlier versions stay readable without migration. On a 10,918-document collection of ~30 KiB documents: orderBy field limit 20 642 → 106 ms, count() 737 → 91 ms, __name__ in ten ids 646 → 10 ms.

REST ListDocuments for a named collection reads the scoped collection index instead of scanning the whole database (#37): pageSize=1 on a 211,202-document store 1.3 s → 1–4 ms.

Fresh-start seed import runs inside one redb transaction with pipelined export decoding and concurrent Storage object copies (#36); Storage import of 33,353 objects 11.1 s → 3.9 s on a Mac. Persistent resume is unchanged.