adapters: model for agents that store sessions in a single database

[fosskar]

If a future target agent keeps all sessions in one SQLite/DB file instead of per-session files, its adapter needs a way to extract/identify individual sessions (e.g. row-level) — the current model assumes one file per session. (This concerns the target agent's storage; ssync's own storage is iroh-docs + iroh-blobs.)

Prerequisite investigation for any of the pending adapters whose agent turns out to be DB-backed.

Refs: TODO.md Open #13.

[fosskar]

First concrete case landed: OpenCode (#8). Schema from sst/opencode packages/core/src/database/schema.gen.ts: single sqlite db at ~/.local/share/opencode/opencode.db (WAL mode), session table (id, project_id, directory, title, time_created/updated, ...) with transcript rows in message/part — row-mutable, no per-session file. Any extraction model must handle: (a) live WAL db owned by a running agent, (b) row-level mutability (no append-only guarantee), (c) reassembling one session = session row + ordered message/part rows.

Contrast worth noting: Codex looked DB-backed at first glance (state_5.sqlite, threads table) but the sqlite is only a metadata index — canonical storage stays per-session JSONL, and codex backfills its index from the files on startup (openai/codex codex-rs/thread-store/src/local/mod.rs). That backfill pattern is exactly what makes file-level sync compose with a DB index; OpenCode has no such file layer. An OpenCode adapter therefore needs genuine row extraction, not a file watcher.