webui-rs 7: Adopt Turso embedded as the local run store (Python writer ↔ Rust reader on one SQLite-format file)

[GeorgePearse]

Extends the Rust + Datastar webui rewrite (#26–#31). Decision 2026-08-05: use Turso embedded (the in-process Rust SQLite rewrite, turso crate — not libSQL-the-fork, not Turso Cloud) as the local run store.

Why

• ProgramDatabase (genesis/database/dbase.py) currently writes runs to ClickHouse via clickhouse_connect — a heavy external dependency for a single-user local research loop (docker-compose, URL env vars, connection health in the webui).
• Turso embedded is in-process, SQLite-file-compatible, async, with native vector search — which maps directly onto the genotype/solution embeddings genesis stores for novelty (EmbeddingClient is already wired into dbase.py). Nearest-neighbour novelty checks become a DB query instead of numpy scans.
• One file per experiment = trivially portable/archivable runs.

Shape

• Writer (Python): ProgramDatabase gains a Turso backend via the official Python bindings (C FFI over the same engine). Schema: programs, generations, islands, lineage edges, embeddings (vector column).
• Reader (Rust webui): the webui-rs 2/6: Port the read model — ClickHouse queries + typed rows in Rust, parity-tested against the FastAPI responses #27 read model queries the same file in-process with the turso crate — no server between the webui and the data at all.
• ClickHouse demoted to optional: keep it as an opt-in sink for shared/long-term history, or drop it outright — decide here and record it. The webui's primary read path becomes Turso.

Risks to burn down first (spike before porting the schema)

1. Multi-process access: evolution loop (Python) writing while the webui (Rust) reads the same file. SQLite solves this with WAL + file locks; verify Turso beta's cross-process story explicitly. Fallbacks if it's not there yet: webui opens read-only snapshots, or the file format's SQLite compatibility means the Rust side can read with rusqlite until Turso matures.
2. Beta API drift: Turso Database is beta (v0.6.x as of May 2026, API still moving; libSQL is the production-ready sibling). Pin versions on both sides; keep the storage layer behind a trait/interface so a swap to libSQL is a contained change if beta bites.
3. Vector search parity: confirm the vector type + ANN query path exists in both the Rust crate and Python bindings before committing the novelty path to it.

Sequencing

Slots in alongside #27: land the spike (risks above) first, then #27's read model targets Turso rather than ClickHouse. Resolves #27's open "legacy sqlite keep-or-drop" question — legacy sqlite mode is superseded by this.