Keep Pixman as the required live-card rendering path

Task 4.3 of the-shell-manages-apps-as-cards asks for a decision on its optional VG-Lite path. The decision is not to pursue VG-Lite as part of this bounded card acceptance path. Continue the required live-card interaction and cost work on Pixman. The separately authorized GPU composition trial remains open with all its source, synchronization, isolation, performance and physical control requirements intact. Nothing in this decision archives that trial or claims GPU acceptance. The machine-readable decision is renderer-decision.json.

The three committed GPU scene comparisons used a retained context and corrected RGB565 upload. GPU passes averaged 43.959–44.452 ms wall time and 37.046–37.259 ms process CPU; paired forced-Pixman passes averaged 36.242–36.647 ms and 32.471–32.646 ms. The result fields in the JSON are copied from the six committed summary files, whose paths are included. These were animated full-panel parent/child scenes, not live-card workloads. They do not justify projecting either a card speedup or an identical card cost. They do identify source snapshots and conversion/upload as unresolved GPU costs.

The latest card comparison retains failed Pixman CPU/tracking budgets despite complete injected interaction traces. Its negative cache result does not establish that GPU would pass. There is no same-workload GPU card measurement, and compositor-only normal-service access and broader cache/operation correctness remain unproved on the physical board. Connecting this optional renderer now would add dependencies without current proof that they solve the measured card gap.

This decision closes only the choice in card task 4.3. Task 4.2 still requires an accepted cost decision that preserves live cards, direct drag, selection, expansion and recoverable throw. Image integration and real-finger acceptance remain open. The user separately authorized making proven RVV support part of the normal image; that work continues under its own published proposal and is not rejected by this renderer decision.

Reconsider a GPU card route when the separate trial establishes usable normal service access, the needed format/clip/scale/completion behavior, and a measured reason to expect the same live workload to improve. Any such comparison must use the card workload and unchanged acceptance limits, retain failed trials, and record physical evidence. Until then, keep the unsupported diagnostic opt-in and do not make it a prerequisite for the gesture shell.

This is a reviewed decision from existing committed physical evidence, not a new hardware test. No source, kernel, image, board state, or acceptance budget was changed. Narrow validation: openspec validate the-shell-manages-apps-as-cards --strict, with referenced JSON values and paths checked against the six source summaries. Physical gates remain explicitly unchecked.