fix(screenshot-diff): compare saved screenshots on any platform

[filip131311]

Fixes #635.

Before / after

$ argent run screenshot-diff --udid chromium-cdp-9222 --baselinePath A.png --currentPath B.png
Tool 'screenshot-diff' is not supported on chromium app (no chromium support declared).

$ argent run screenshot-diff --udid chromium-cdp-9222 --baselinePath A.png --currentPath B.png
{ "summary": "...status: unchanged...", "diffPath": "...", "contextDiffPath": "..." }

Comparing two saved PNGs touches no device, yet the tool refused by platform — so visual regression
testing was unavailable for Electron entirely. Two platforms the issue doesn't mention were refused
the same way: remote iOS (remote:*) and Vega.

Why the capability is now shared rather than widened

The list was a second copy of the one screenshot keeps, and they had drifted: a screenshot can be
captured from five platforms, a diff was allowed on two. It is the same question — this tool's only
device interaction is a screenshot — so they are now one object. That is what stops it drifting
again, and it is why there's no hand-written platform allow-list anywhere in this diff.

The narrower thing is live capture, so that is where the limit sits now: requesting one on a
platform whose screenshots come from a backend this tool doesn't drive says so and points at the
two-file form. Compare the old error, which is what the issue objected to:

before: Tool 'SimulatorServer' is not supported on ios simulator (this is an Apple TV (tvOS)
        simulator — touch, paste and rotate input are not available...)      ← never says "screenshot"
after:  Cannot capture a screenshot for a diff on chromium. Capture it separately with the
        `screenshot` tool, then pass both files as baselinePath and currentPath — comparing saved
        images works on every platform.

Those platforms also stop requesting the simulator-server, which would otherwise fail while
resolving the service — before any of this tool's code runs — which is exactly how the caller ended
up seeing a blueprint error instead of one about screenshots.

What I deliberately did NOT do

An earlier draft extracted screenshot's capture into a shared helper so live capture would work on
Chromium/tvOS/Vega too, and made services() return {} unconditionally. I dropped it:

• It would have silently removed the simulator-server self-heal. _recoverFailedServices iterates
  the refs from services(), so returning {} makes recovery unreachable — and
  simulator-server.ts:293-295 documents that the wedged-server case emits no exit, so that retry
  is the only recovery. A user rebooting a simulator would have gone from a transparent retry to a
  manual stop-simulator-server. That is a regression on this tool's primary documented flow, traded
  for a file↔file fix.
• rotation would have been silently dropped on tvOS and Vega (neither capture function takes it)
  and on Chromium without the optional sharp. For screenshot that means a wrongly-oriented image;
  for a diff it means a wrong verdict against a rotated baseline.

Live capture on those platforms is worth doing — as its own PR, with rotation, abort-signal and
topMask handled deliberately. This PR keeps screenshot untouched.

Also noted, not fixed here

topMask defaults to "status-bar" and blanks the top 6%. On an Electron window or a TV screen there
is no status bar, so a real change in that band is ignored. It is not exposed in the schema. Now that
those platforms can diff, it's worth a follow-up — but changing it would alter the already-working
path, so I left it.

Verification

Live, against the built CLI:

	before	after
chromium, two saved PNGs	refused	real diff
chromium + captureCurrent	refused as unsupported tool	error naming screenshots + the workaround
tvOS, two saved PNGs	worked	still works
vega, two saved PNGs	refused (unreported)	real diff

tool-server 3096 passed. The existing screenshot-diff-tool.test.ts passes unmodified, including
its guard that the simulator-server is declared only for live captures — that guarantee is preserved
rather than deleted. New tests assert at the capability gate (assertSupported), because that is
where the refusal happened; a test driving execute cannot see this bug at all.

🤖 Generated with Claude Code