fix(cu): preserve unknown dispatch outcomes

[hqhq1025]

Summary

Follow-up to the non-blocking Runtime finding on merged #911.

• preserve typed backend failures such as outcome_unknown after dispatch
• keep the session fenced in reobserve_required
• run post-dispatch lease rejection only for backend successes
• cover both generic coordinate and semantic element action paths

Root cause

applyTypedOutcomeState() correctly advanced the session after outcome_unknown, but the subsequent old-lease validation replaced the backend result with reobserve_required. The state was safe, but the model lost the critical signal that delivery may already have occurred and could retry an action incorrectly.

Verification

• Runtime typecheck
• focused Computer Use/session-state suite: 69/69
• git diff --check

[hqhq1025]

@Astro-Han Follow-up to your merged #911 P2 is isolated and CI typecheck/test/e2e is green. Both generic left_click and semantic click_element now preserve model-visible outcome_unknown while the session remains fenced in reobserve_required. Ready for review.

[Astro-Han]

Approved. I did not find a P0 or P1. This is the follow-up round for #911. One non-blocking P2 remains; I did not find an additional P3.

[P2] Revalidate host-read leases even when the backend fails. cursor_position and wait use the observation lease, but gating that check on result.outcome.ok lets a failed host result complete after clearSession(). I reproduced both calls returning service_unavailable after the session had already become user_stopped. This regresses the stop fence added in #911; the existing test covers only successful late results. Keep observation-lease validation unconditional, and narrow the outcome_unknown exception to delivered action results whose session was not changed externally.