Add semantic act-wait-observe execution

[wolfiesch]

Why

Agents currently have to combine brittle node references, fixed sleeps, and repeated snapshots. That is slower, less ergonomic, and particularly unreliable when the user leaves a task unattended.

What changes

• Adds stable semantic accessibility refs for unique actionable role + accessible name targets.
• Re-resolves a semantic ref against the live accessibility tree so SPA DOM replacement does not require a new snapshot.
• Fails closed with actionable candidates when a target becomes ambiguous or disappears.
• Binds resolution to the exact live document identity and page revision before and after the AX query, preventing a stale ref from crossing a navigation/origin boundary.
• Adds event/hybrid waits for load, URL, DOM text/selector, network idle, and download completion.
• Anchors network-idle quiet time to each wait and wakes once an active request burst drains.
• Filters irrelevant tab metadata events and coalesces MutationObserver scans to bound page-churn overhead.
• Adds TypeScript/Python actWaitObserve helpers that stop correctly on Commit/Your Turn outcomes.
• Enables optional debugger domains lazily, recovers from domain-disable/failure, and removes global observation as a barrier to later independent writes.

Review decisions

• Semantic identity: refs intentionally encode revision, role, and name—not CSS structure. Duplicate matches fail closed instead of guessing.
• Wait semantics: network idle uses a 500 ms quiet window and DOM identity is revalidated; these are practical heuristics rather than a transaction with the page.
• Helper atomicity: act → wait → observe is a documented multi-RPC workflow. Another actor/page script may still change state between phases.
• Ordering: a global observation waits for earlier writes but no longer blocks unrelated writes submitted afterward.

Validation

• Workspace TypeScript typecheck/build and extension build
• Semantic uniqueness, SPA replacement, ambiguity, and cross-document race regressions
• Load/status/URL wake and 100-event storm regressions
• DOM observer coalescing/options regressions
• Network-idle anchoring/burst tests
• Lazy debugger-domain and Developer-disable recovery tests
• TypeScript and Python helper outcome tests

Why draft

This is a substantial execution-model and scheduling change. It should be evaluated with representative real sites and agent traces before it becomes the default recommended workflow.