Carve the engine into zigstore v1.0.0; consume it as a dependency

[poly-glot]

Extracts dmozdb's generic DB/server engine into the standalone zigstore library (pinned at the v1.0.0 tarball + content hash) and rewires dmozdb to consume it. Big-bang, landed as one validated PR per docs/superpowers/specs/2026-06-06-zigstore-library-extraction-design.md and its implementation plan.

What changed

• database.zig → app Directory composing zigstore.Engine(schema)'s *Store; the 11 indexes / 6 memtables / 3 counters are now a comptime schema.
• types.zig split into zigstore.codec + app schema.zig; changeset codec → zigstore.wire_codec; binary-protocol framing → zigstore.protocol; epoll/bootstrap → zigstore.run; gen_client_ts → a zigstore.tsgen driver.
• 18 moved engine files deleted; dmozdb now reaches the engine only via @import("zigstore") (one-way boundary, verified).
• Engine→app couplings inverted into runtime seams: recover hooks, spawnWorker ticks, the commit ctx callback, processFrames dispatch, SnapshotHost.

Behavior preserved

• 193/193 tests green; the binary-protocol wire format and the generated TS client (web/lib/dmoz-protocol.gen.ts) are byte-identical; the server boots end to end (bootstrap Top=1, 5 epoll reactors).
• The pipeline RESPONSE_RESERVE is now an app-supplied worst-case bound (restored from the regressed 64 KiB hardcode), with a regression test.

⚠️ Before deploy

• The on-disk format changed (store.dat, schema-generated header) — production needs a data re-seed via the importer.
• Run the Fresh + Playwright + visual check in the devcontainer (per the project contract).

Follow-ups (non-blocking)

• Internal cleanups: add an engine Store.drainMemtable(name) seam so the app stops duplicating drainOneInner; move WAL truncate-after-checkpoint into Store.deinit to drop the shutdown double drain/flush.