feat(snapshot): disclose a cut capture on every platform; raise the iOS bridge node cap to 5000

[thymikee]

Part of #2492 (task 5). Two changes, both measured first.

One disclosure for a cut capture, every platform

Every backend already reports the same fact — truncated: true on the snapshot — but only in JSON. The text path said nothing: an agent read a screen missing its footer, tab bar, or the items after a long list as complete. The backends walk the tree in document order, so what falls off is what comes last, on screen or not.

truncatedCaptureWarning in capture-kit renders one warning from that flag, and the single cross-platform warnings assembly in the snapshot runtime calls it. Android helper (5000-node cap), iOS Simulator bridge, XCTest runner, and the web provider all flow through it. The limit and dimension stay backend-side; the copy names neither, and tells the agent what to do: navigate or scroll so fewer elements render, re-run, use screenshot for the rest.

Producers that cannot observe truncation keep truncated unknown (#2188 invariant 5) and get no warning.

iOS Simulator bridge node cap 1500 → 5000

Measured on iPhone 17 Pro with a synthetic unvirtualized list in the fixture dev client (8 regular snapshots per cell, bridge acquire duration from --debug diagnostics):

tree	cap 1500	3000	5000	10000
600 rows (~3.5k raw nodes)	969 ms, truncated	976 ms, truncated	962 ms, complete	970 ms, complete
1500 rows (~8k raw nodes)	2378 ms, truncated	–	2344 ms, truncated	2404 ms, complete

The native read fetches the whole tree regardless of the cap; the cap only stops conversion. Cost is the tree itself (~0.25 ms per raw node). At 1500 the screen's on-screen footer (drawn last) was absent; at 5000 present. 5000 matches the Android helper. The 4 MB response bound did not bite at 8k nodes.

Not here

• Falling back to XCTest for a capture the bridge still cuts at 5000: a policy decision (iOS Simulator bridge and web content: follow-ups to the #2484 fence (content-class gate, probe circuit, WebContent spike) #2492 task 5), not a performance one — such a screen already costs over a second per capture.
• Acquisition-rooted --scope on the bridge as a real override; today --scope is presentation-only on iOS, so the warning does not promise it.

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://callstack.github.io/agent-device/pr-preview/pr-2510/
Built to branch gh-pages at 2026-09-12 05:58 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.50 MB	4.50 MB	+407 B
Package (unpacked)	4.50 MB	4.50 MB	+407 B
Package (download)	1.32 MB	1.32 MB	+162 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.2 ms	26.5 ms	+0.3 ms
CLI --help	74.0 ms	73.6 ms	-0.4 ms

[thymikee]

The new warning misclassifies incomplete captures at a552b16. quality-warnings.ts:43-47 treats every truncated flag as a backend-limit cut, but Android presentation failures and iOS accessibility-unavailable responses also set that flag. Those failures now get misleading advice about missing later elements and scrolling. Use wording that describes incomplete capture without inventing its cause, or select limit-specific advice from typed evidence; cover a real sparse/presentation-failure result. Current-head checks pass and there are no conflicts. The separate bridge publish-versus-fallback policy remains deferred, as the description states.