test(rivetkit): cover failed manual startup initialization

[NathanFlurry]

• Add regression coverage for manual startup failures before initialization is durably marked complete.
• Verify retry startup still sees the actor as uninitialized and only commits the flag after success.

[claude]

Review

This is a test-only PR adding regression coverage for a real correctness invariant: a manual-startup runtime that fails before signaling startup_ready must not leave has_initialized = true durably persisted, so a retry is correctly treated as a fresh (is_new) startup.

Verified against the actual implementation (rivetkit-rust/packages/rivetkit-core/src/actor/task.rs::start_actor/spawn_run_handle):

• For a new actor with requires_manual_startup_ready() == true, the CoreInit phase intentionally skips set_has_initialized(true) + persist_state(...) (line ~1183), deferring to the runtime preamble.
• If the manual runtime's startup_ready reply is Err(...), spawn_run_handle propagates the error via .context("runtime startup preamble")? before start_actor ever reaches the PostReady fallback block that would set has_initialized.
• So the test's assertion that maybe_load_persisted_actor(&ctx).await.is_none() after a failed start, followed by the retry observing has_initialized == false and only persisting true after a successful manual Ok(()) reply, is an accurate exercise of this code path. Good regression coverage for a bug class that would otherwise silently corrupt the "is this actor new" determination on retry.

Style/consistency:

• The new test closely mirrors the existing sibling test directly above it (manual_startup_does_not_mark_initialized_before_runtime_preamble), reusing the same std::sync::Mutex-wrapped oneshot pattern for observing has_initialized inside the closure, and the same shared-kv fixture trick (kv.clone() then reusing kv) to give both ActorContexts the same underlying in-memory sqlite fixture (confirmed via Kv::test_identity() returning a stable Arc::as_ptr, so the fixture key matches across the two contexts). This is consistent with the file's established conventions.
• maybe_load_persisted_actor is a small, appropriately-scoped addition alongside the existing load_persisted_actor helper.
• Correctly follows the repo's "Rust tests live under tests/" convention — this is already in tests/task.rs.

Minor nit (non-blocking): the first (task) run handle from the failing manual-startup attempt is never explicitly .abort()/awaited, unlike retry_task's at the end. This is harmless here since the closure sends its error and returns immediately (no long-lived loop), but for consistency with the rest of the file's tests you could join/abort it too.

No production code is touched, so there's no functional, performance, or security risk from this change. Nice, tightly-scoped regression test.