feat(online): end-to-end handshake between game client and local server

[aram-devdocs]

The plugin running inside Sailwind connects to the local Rust server, completes a versioned handshake including the API surface hash, and round-trips a position message.

Blocked by #13 and #14. The wire path is proven by protocol-smoke (7/7, real LiteNetLib client vs Rust server); the in-game leg is verified in Phase F.

[aram-devdocs]

Partial (init-0). The full handshake path is proven by protocol-smoke: real LiteNetLib client -> Rust server, ClientHello/ServerHello with protocol_version and a position round-trip, plus SQLite restart persistence and server-driven AoI (7/7, green in cloud CI too). Remaining: the in-game leg — the plugin inside Sailwind completing the same handshake including the live API surface hash (ties to #5).

[aram-devdocs]

Handoff: The wire handshake is proven headless — make smoke 8/8 incl. checks 1–2 (real LiteNetLib client ↔ Rust server, ServerHello accepted with the API surface hash). Remaining (needs a save loaded): make run-modded → load a save → confirm the in-game ClientHello→ServerHello + a boat-position round-trip in the BepInEx log.

[aram-devdocs]

Picking this up. Branch: feat/15-ingame-handshake

[aram-devdocs]

Headless re-verification on eat/15-ingame-handshake is green: the Release Rust server built and dotnet run --project tools/protocol-smoke passed 8/8. That includes a real LiteNetLib ClientHello/ServerHello exchange (protocol v1) and the position path: client B observed client A's added cell plus 8 monotonically increasing X samples after A streamed ClientState updates. The remaining acceptance boundary is the in-game plugin leg with the live Sailwind.API surface hash. That leg depends on the generated-token/world-readiness fix in #44 (Fixes #5), which is open and fully green. Runtime launch/testing is intentionally deferred while the game is in use; no Steam/Sailwind process was started.

[aram-devdocs]

Live game verification is complete using disposable slot 2. Both Continue and New Game reached gameplay and connected to the isolated server; the original New Game loading hang did not reproduce. A second LiteNetLib client also verified bidirectional position replication with two distinct samples. All protected slot-0 save hashes still match the pre-test baseline. Full evidence is recorded on PR #47.