feat: persist request history so Transactions survives a restart

[flyworker]

The Transactions view read a 1000-entry in-memory ring, so every restart emptied it. You hit this directly: after a deploy the list showed 17 rows and total_requests read 3.

This node restarts often enough that the earnings panel routinely reports counter resets, and each one wiped the list — which also meant the source filter and pager from #118 could only ever page through traffic since the last restart.

What changed

Each served request is written to a request_history table beside the existing metrics_history, and /inference/requests reads from it. The in-memory ring stays as a fallback and as what the model detail view reads.

Design constraints worth reviewing

Recording must never slow down serving. The database allows a single open connection (SetMaxOpenConns(1)), so a synchronous insert per request would serialise inference behind SQLite. Writes are queued and batched by a background writer; a queue full enough to overflow drops records and counts them rather than blocking. Shutdown drains what is queued.

One row per request, so retention has two limits. RetentionDays (default 7) is how far back it stays useful; MaxRows (default 200000) stops a burst filling the disk before the age limit ever applies. Whichever binds first wins; pruning runs hourly and on start. Error reasons are truncated to 512 bytes for the same reason — an upstream body is unbounded and there is one per row.

Unlabelled rows count as Hub. Rows written before the source column carry none, and all arrived over the WebSocket. Filtering to Hub includes them, matching the in-memory path — otherwise older history vanishes the moment you touch the filter.

On SQLite and data files

SQLite is already a pure-Go dependency (glebarez/sqlite → modernc.org/sqlite), so it is compiled into the binary and works under the CGO_ENABLED=0 cross-compiles — nothing needed there.

No data file was ever tracked, but .gitignore had no *.db pattern and the docs say CP_PATH can point at the repo. Now ignores *.db, *.db-wal, *.db-shm, *.sqlite, *.sqlite3.

Verified

8 new tests against a real SQLite database, not a stand-in: paging and newest-first ordering, model and source filters, unlabelled rows counting as hub, history surviving a simulated restart (a second store over the same file sees the first one's writes), pruning by age, pruning by row cap keeping the newest, Record never blocking when the queue is full, and long error truncation.

8/8 packages pass, go vet and gofmt clean. Config documented in docs/configuration.md.