fix(cu): fence failed host reads after stop

[hqhq1025]

Summary

Follow-up to the remaining P2 on merged #921.

• revalidate observation leases after cursor_position and wait regardless of backend success or failure
• discard late service_unavailable results after clearSession()
• preserve the action-only outcome_unknown exception from fix(cu): preserve unknown dispatch outcomes #921

Verification

• Runtime typecheck
• focused Computer Use/session-state suite
• regression coverage for failed cursor and wait calls completing after stop
• git diff --check

[hqhq1025]

@Astro-Han Follow-up to the remaining merged #921 P2 is isolated and CI is fully green. Failed cursor_position/wait results that complete after clearSession are now discarded as user_stopped; the action-only outcome_unknown exception remains intact. Ready for review.

[Astro-Han]

No P0/P1 findings. One non-blocking issue remains:

• P2: A normal cursor_position or wait failure such as service_unavailable is converted to reobserve_required. applyTypedOutcomeState() changes the generation before the new observation-lease validation runs, so the original typed error is lost even when no stop occurred.

Checking the pre-existing lease before applying the returned outcome state would preserve both behaviors: late results after clearSession() remain user_stopped, while ordinary backend failures retain their typed error.