fix(sidecar): stop wasm stdin stream-event flood; tolerate shadow-sync races

[NathanFlurry]

• Sidecar-managed wasm guests read stdin through the kernel (__kernel_stdin_read), but every host stdin write also emitted a stdin stream event to the guest's V8 session. Nothing consumes those events for wasm, and while the guest thread is blocked in a sync bridge call each one lands in the session's bounded deferred-message queue — one dead event per keystroke, killing the session with "sync bridge deferred message queue exceeded limit of 256" after ~256 keystrokes in an interactive shell. Wasm stdin writes now feed the kernel only.
• The host-shadow → guest reconciliation walk failed the entire VM when an entry vanished mid-walk or the guest-side path could not receive the write (editor swap-file churn, e.g. vim's .swp/.swo dance). One transient race then poisoned every subsequent fs op on the VM (ls returning i/o errors, vim E212 on write, sessions dying on the next sync). NotFound/ENOENT during the walk is now treated as "entry mid-churn, skip", matching native semantics.