fix(Alden): distinguish stopped browser and pre-press AX actions from failures

[twoimo]

Precise emergency-stop attribution for Alden's owned browser and opt-in background Accessibility tool runtime. Stacked on draft PR #31 (which is stacked on #30→#29→#28→#27). No existing branch is merged or forced.

• Browser runner failure after the global abort latch becomes ABORTED / AX_ABORT_GLOBAL with no leaked exception/task/result text. A non-aborted exception remains a bounded browser failure. No claim that previous third-party browser effects were rolled back.
• Exact AXPress tool returns the actual cancellation only if it occurred while resolving the target before the press, even when resolver throws FocusStealRequired, target resolution error or provider exception. No AXPress occurs after that state.
• Deliberately preserve the existing AX_ERROR_EFFECT_UNKNOWN contract for an action after the press may have happened; an uncertain effect cannot be reclassified as safely cancelled, or automatically retried.
• Retain exact PID/window/role/identifier, explicit opt-in, background no-focus/no-real-pointer enforcement, bound output and global abort epoch contracts.

Test evidence: macOS Actions #38072772583, source SHA ef94d51226314408fbcd8830f95d2728bb783cea, 92/92 Python regression tests passing (browser ownership, AX, cancellation, input fences and underlying voice-turn path). Production and test Git blobs match the implementation branch; runner/stdiotoken mocks cannot establish physical Mac Accessibility outcomes.

Technical design/rollback requirements.

Not performed: Mac host access, direct original goal-objective or user-local outputs/STATUS.md read/write (native Core requires an unprovided valid turn_token), production DB/model/queue/process changes, integration with external osk-system, signed/notarized distribution, installation or user hardware controls.