Move the eager state map into the invocation states

[jackyzha0]

Problem

CoreVM keeps the invocation's eager state map (Context.eager_state, the complete Virtual Object KV state) alive until the CoreVM value is dropped. On native SDKs that's fine — the VM is dropped promptly at end of invocation. In the WASM SDKs it leaks by default: nothing calls WasmVM::free(), so the wrapper is reclaimed only when the host garbage collector's FinalizationRegistry fires. Under load those finalizers lag, completed invocations' state maps accumulate inside the shared WebAssembly.Memory, and since wasm linear memory never shrinks, the high-water mark becomes a permanent external-memory floor for the process.

We hit this in production with @restatedev/restate-sdk@1.16.4 (Node, Virtual Objects with multi-MB state maps): a flat ~319 MiB (external) ArrayBuffer plateau out of ~615 MiB live heap, alongside GC-overhead and CPU-starvation signals. wasm32 also caps the shared linear memory at 4 GiB, and allocator exhaustion there fails every in-flight invocation at once, so host memory sizing can't fully mitigate.

Why this shape

The state machine already has the right drop mechanism: Transition consumes self by value, which is why Replaying's command buffer and Processing's async results never had this problem. The eager state map was simply owned by Context, which outlives the machine. This PR moves it into the states that can observe it (WaitingReplayEntries/Replaying/Processing), so State::Closed structurally cannot retain it and failed transitions drop it automatically — no release hook, no terminal-edge analysis, no host-language lifecycle coordination.

Changes

• eager_state moves from Context into the three non-terminal State variants.
• State writes get dedicated SysStateSet/SysStateClear/SysStateClearAll transitions (mirroring SysStateGet), which journal the command and then update the map inside the machine, following the delegate-then-mutate shape of SysNonCompletableEntryWithCompletion. This also fixes a small pre-existing quirk: the map used to be mutated before the state check, so a write in an invalid state mutated it and then errored — it now journals nothing and mutates nothing.
• The replay helpers borrow commands/async_results and return the notification handle; the triplicated continue-replaying-or-start-processing construction is deduplicated into replaying_or_processing().

Behavior

Unobservable to user code: reads consult the map only inside the Processing arm, post-terminal sys_* calls already fail with ClosedError/UnexpectedStateError before reaching it, and take_output is untouched. The existing test suite passes unchanged.

Happy to follow up with the sdk-typescript dependency bump + wasm bindings regeneration, and with before/after production memory numbers.

🤖 Generated with Claude Code

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[jackyzha0]

I have read the CLA Document and I hereby sign the CLA

[slinkydeveloper]

Thanks for the contribution!