v2.8.2

Targeting misses now say why

A tap or input_text miss covered three different situations with one message: the target was never there, someone drove the app in parallel between the agent's plan and the call landing, or the matcher was wrong. All three were counted as validation failures — which is why a 25.8% input_text failure rate said nothing about whether input_text works.

• Parallel interference is now named. The screen is compared against the agent's last look at it, with changes attributed to the agent's own tools first. When nobody's tool moved it, the error says so and tells the agent to re-read the screen instead of guessing at the predicate.
• "showing 1 of 4" no longer lies. A complete list of matching inputs was rendered as if it were truncated, which is what sent agents guessing at a higher index.
• An unverifiable write is no longer a clean success. input_text reports whether it wrote and whether the text landed as separate facts, and captures a failure artifact when either goes wrong.
• Identity memoisation for the hardware machine id lookup.