feat(desktop): snapshot diff — return only what changed

[rrader26]

Summary

• New tool `agentmark_desktop_diff`: take a fresh capture, return the structured delta vs the last cached snapshot.
• New library function `diffDesktopCaptures(before, after)`: pure, exportable, reusable.
• `convertDesktop()` now returns the raw `DesktopCapture` via the new `DesktopConversionResult` type so the session can cache it for the next diff.

Why

Smaller token payload + faster reasoning for two extremely common agent questions:

• "Did my action take effect?" → check `elements_changed` for the field that should have moved.
• "Did a new dialog appear?" → check `elements_added` for a `dialog`-role node.

Today an agent answering either of those has to re-snapshot the full tree and reason over thousands of tokens. With diff, "no_changes" is one boolean; even a busy diff is typically <500 tokens.

Also: prerequisite for Recipes (the record/replay feature next on the roadmap). Recipes use diff as the verification primitive after each replayed step.

Diff shape

```json
{
"no_changes": false,
"summary": { "added": 1, "removed": 0, "changed": 2 },
"window_title_changed": null,
"focus_changed": { "from": "btn_save", "to": "txt_name" },
"elements_added": [{ "id": "...", "role": "...", "name": "...", "value": "..." }],
"elements_removed": [{ "id": "...", "role": "...", "name": "...", "value": "..." }],
"elements_changed": [
{ "id": "...", "role": "...", "name": "...", "changes": {
"value": { "from": "old", "to": "new" },
"aria.checked": { "from": false, "to": true }
}}
]
}
```

Semantics

• Matching is by stable accessibility id (`DesktopElement.id`). Roles + names are auxiliary, not used for matching.
• Sub-pixel bounds jitter (≤1px) is suppressed — compositors produce that constantly and it's never actionable.
• After a successful diff, the session's cached capture moves forward. The NEXT diff is against this new baseline, not the original snapshot.
• The action-id binding is not updated. Call `agentmark_desktop_snapshot` when you need fresh action_ids after a structural change. Diff is read-only on actionable state.

Test plan

• `pnpm build` clean
• `pnpm test` — 354 pass / 10 skip (7 new: pure-function coverage + tool integration)
• Manual: type into a NowCerts field, diff, confirm only that field appears in elements_changed

Roadmap dependency chain

This unblocks the Recipes PR, which uses diff to verify each replayed step ("did the expected change happen? if not, fail this recipe step instead of plowing through stale state").

🤖 Generated with Claude Code