Commit 4afca05 switched from page.frames() to mainFrame().childFrames()
on the assumption that processFrame recurses. It doesn't — processFrame
only serializes and returns. As a result, topologies like
main(A) -> same-origin(A) -> cross-origin(B) silently dropped B.

Option A (chosen, minimal): walk the full page.frames() tree and skip
frames whose ancestor is already a known cross-origin frame. This
preserves correctness for all observed topologies while keeping each
top-level CORS frame processed exactly once (descendants are stitched
in by the in-frame PercyDOM.serialize call).

Picked over Option B (recursive processFrame) because the diff is
smaller and the existing test suite's mocking around mockPage.frames()
keeps working as-is.

Also leaves a TODO in processFrame for closed-shadow exposure inside
CORS iframes — that requires a CDP session attached to the OOPIF
target, which is materially more than ~30 LOC and is tracked separately.
Top-page closed-shadow capture is unaffected.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

A single detached backendNodeId (host or shadow) caused DOM.resolveNode
to throw, which previously aborted the whole loop and lost every
remaining closed shadow root on the page. Wrap each iteration in its
own try/catch so one bad pair only skips itself.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Covers the three regressions guarded by the BLOCKER fix and the
per-host try/catch:

- crossOriginGrandchildOfSameOriginFrameIsStillCaptured
  main(A) -> same-origin(A) -> cross-origin(B): B is emitted.
- descendantsOfACorsFrameAreNotRecapturedIndependently
  main(A) -> cors(B) -> inner(B) -> nested-cors(C): only B is emitted;
  C is handled inside B's PercyDOM.serialize.
- domEnableIsPairedWithDomDisableOnSuccess / OnException
  DOM.disable always runs after a successful DOM.enable, even when
  DOM.getDocument throws. The CDP session is always detached.
- domDisableNotCalledWhenDomEnableFails
  If DOM.enable itself throws, DOM.disable is not called (correct
  protocol shape) but detach still runs.
- oneBadBackendNodeIdDoesNotAbortRemainingShadowHosts
  One DOM.resolveNode failure skips only its pair; the remaining
  closed shadow hosts are still exposed via Runtime.callFunctionOn.

Tests are pure-mock (Mockito) — no Playwright browser required, so
they run on the full Java 8/11/17/21 CI matrix.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

# Conflicts:
#	package.json

Percy.java (PER-7292) imports com.google.gson.* directly for parsing the
CDP DOM tree when exposing closed shadow roots, but gson was only present
transitively via the Playwright artifact. A direct compile-time source
dependency must be declared explicitly so a future Playwright upgrade that
drops or relocates gson cannot silently break compilation.

Pinned to 2.10.1 to match the version Playwright 1.44.0 currently resolves,
so no version skew is introduced. Build and offline test suites unaffected.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>