No guessing

Two changes, both about refusing to guess: undo will not run on a PC it cannot prove is the right one, and the AI is now physically unable to name a repair that does not exist.

• Undo now proves it is on the right PC before it runs. Every repair snapshot records which PC it was made on, and undo already refused to restore a snapshot from a different PC. But when that identity check itself failed, a rare case where the machine's identity cannot be read at all, the undo went ahead and assumed everything was fine. That assumption is gone. If the app cannot confirm it is on the same PC the snapshot came from, the automatic undo refuses and tells you exactly how to restore by hand instead: the snapshot folder, its undo.ps1 script, and the .reg files it contains perform the same restore. Snapshots from before PC identity was recorded still restore as they always did
• The AI can only ever name a repair that actually exists. The AI has never been allowed to write repair commands; it picks repairs from the app's fixed catalog, and that boundary is the foundation this product is built on. But the decoder that forces the AI's diagnosis output into strict JSON left the repair name field free-form, and in our own measurements the models filled it with invented names constantly: 29 of 100 test diagnosis calls from one shipped model named a repair that does not exist. Those invented names were always discarded before they could do anything, so this was never dangerous, but silently discarding nonsense is still a form of guessing. The list of real repair names is now built into the decoder itself, generated from the same catalog the app executes from, so a made-up name cannot be produced at all. We re-ran the same 100-call measurement after the change: zero invented names
• An honest footnote on that guarantee. It also made the AI more accurate, because the decoder now spends its effort choosing among real repairs instead of inventing fake ones. And to be precise about scope: this decoder is exercised today by our test and evaluation instruments, while everyday scans use deterministic rules and chat answers are plain prose. It is a guarantee about the AI layer itself, inherited by everything built on it, not a visible change in this week's app
• No new repairs in this release, and the module catalog is unchanged