fix(cua-driver): isolate screenshot transforms between clients

[injaneity]

Superseded by issue #3630

At maintainer request, this work has moved to #3630 as a defect/evidence record for architectural review alongside RFC #3473 and active PR #3616. This prototype is closed without merging. Its branch and commits are retained for regression-test reuse and provenance; do not treat the session-registry design or its refusal changes as approved. The validation below is the historical prototype record, not a current readiness claim.

Scope

Refs #1882. This draft fixes cross-client screenshot transform interference, rather than claiming that the metadata/documentation issue is fully resolved. It does not introduce image-token machinery or supersede RFC #3473 / PR #3616.

Current head: 21fad867081f6b69d91418c39572aededdbace22.

Confirmed defect

Two persistent MCP clients captured the same unchanged 1470×932-point disposable AppKit window using installed macOS cua-driver 0.23.2. Client A requested a 200×127 image; client B requested native resolution. A clicked the same image coordinate (99.5, 62.5) throughout:

Sequence	Native receiver landing, window-top coordinates	Outcome
A capture → A click	(731.325, 459.375)	Target
A capture → B native capture → A click	(99.5, 62.5)	Wrong region
A refresh → A click	(731.325, 459.375)	Target

The receiver's mouse events, not the driver's unverifiable action response, are the oracle. Window geometry and backing scale remained unchanged. The baseline binary SHA-256 was 67ccfc99e69ebb5881fdfc3787d85abcd8cc2beb7423f255549557623cab6907; its reported source SHA was null. This is installed-version diagnostic evidence, not exact-main or candidate certification. No existing user browser tabs were targeted.

Implementation

• Shared core resize registry keyed by trusted private lifecycle identity, PID, and window ID.
• macOS, Windows, and Linux captures and resize consumers use that registry, including delegated pixel-focus paths, Linux held-pointer/recording conversion, and Windows zoom crop conversion.
• Native-size captures publish identity transforms instead of clearing another client's scale.
• Session cleanup removes only the owner's transforms; late capture completion cannot resurrect an ended session.
• Common dispatch refuses missing or ambiguous window-pixel context with screenshot_context_missing, rather than borrowing another client's transform or guessing native pixels.
• Documented persistent MCP / explicit CLI session usage. Unnamed one-shot CLI calls cannot retain screenshot context across processes.
• Shared unit and trusted-dispatch regressions, plus same-daemon two-client interference in existing AppKit, WPF, and supported GTK3 pixel rows. Their fixture-state and desktop oracles remain required; the observer connection does not create an unrelated recording.

Validation so far

• cargo test -p cua-driver-core --locked: 620 passed (615 library, 5 integration); one documentation test ignored. Rerun against the final local source before the latest commit.
• cargo test -p platform-macos --lib --locked: 365 passed, 2 ignored, at dfd15de01f00b85d3685aba9039e5529ffcd5acc. The subsequent product diff is a Windows-only zoom call-site correction; remaining changes are tests/testkit/docs.
• cargo test -p cua-driver-testkit --lib --locked: passed after adding the peer-connection helper.
• AppKit native regression target compiles after the peer-connection update; this is not a native execution result.
• Source driver build, formatting, and diff checks passed.
• Previous-head Linux native unit/compile, Linux/macOS portable contract parity, generated contract/SDK checks, and release metadata passed.
• Previous-head Windows CI caught one missed zoom resize lookup; this head supplies session/window identity there. Windows compilation and dependent checks are rerunning, not yet claimed green.

Readiness blockers / boundaries

• Native candidate replay and the canonical stable-SHA desktop matrix. The source-built private daemon's read-only permission check reported Accessibility=false and Screen Recording=false; the separately installed signed application is authorized. No TCC reset, installed-app replacement, or permission prompt was performed. The private probe daemon has been stopped.
• Reconcile and verify Wayland refusal precedence: when a compositor cannot supply a window screenshot, the new common missing-context precondition can precede the adapter's background-unavailable refusal. Existing refusal rows must explicitly cover that ordering before readiness; X11 success is not Wayland certification.
• Latest-head Windows/Linux CI, native harness compilation, and documentation/contract checks.

A later capture in the same session still replaces that session's previous transform. This draft does not claim image-ID freshness, window/display-change invalidation, zoom-context ownership, or every recording-rendering alignment case. screenshot_scale retains its existing native-backing-scale meaning.

The PR remains draft and must not be treated as merge-ready or exact-candidate desktop-certified.

[injaneity]

superseded by #3630 at maintainer request. the issue preserves the reproduction, prototype provenance, validation limits, and architectural follow-up with #3473 / #3616. closing without merging or deleting the branch; no expansion of the selected rfc scope is implied.