fix(present): no-vsync fullscreen present flood + fullscreen direct scanout (Phase A + Phase B)

[ariel3259]

Summary

A no-vsync fullscreen game (CS2) made the server stutter: page_flip/s collapsed from refresh (~60) to 27–47 Hz because every PresentOptionAsync present re-composed the full 1920×1080 scene. This PR fixes the present flood at the core and enables fullscreen direct scanout (Phase A + Phase B).

What changed

Phase A — async present defer + supersession (core)

• Park async presents while a flip is in flight (22674c7e): a PresentOptionAsync arriving mid-flip no longer re-composes; it parks.
• Async successors supersede parked async predecessors (ca88a14b): a newer async present scraps the parked one (Xorg scrap semantics), so the flood coalesces to roughly one present per flip.
• Superseded async victims get their Skip with eff=0 (225551f4): immediate gate, outside the synced hold-back.
• Synced-present behavior is bit-for-bit unchanged.

Phase B — fullscreen direct scanout (KMS)

• Unredirected fullscreen windows are the authoritative root (6182aa52): their presents become direct-scanout eligible.
• Fullscreen explicit-sync presents accepted on the direct path (37bd6c0c).
• Fullscreen sources are pre-probed before admission (2d7ab515), with the atomic TEST_ONLY m1 probe.
• Degrade failed composed unflip instead of exiting the server (4c98c3fd): if the synchronized atomic unflip fails, fall back to per-output composed flips without releasing the still-scanned dma-buf or killing the session.
• Late-materialization store ref (e10f4cbd): a deferred store ref is recorded for Pictures whose backing materializes late, fixing the game-start transparency bug where free_pixmap destroyed a drawable under a live Picture.
• cargo +nightly fmt across the Phase A/B changes (3fa1ec7c).

Key design decisions

1. options=0x8 is NOT async. The original finding assumed PresentOptionAsync; 0x8 is PresentOptionSuboptimal (synced). The observable fix on the nvidia box comes from the pre-existing synced supersession + the merged DRI3 syncobj fixes (PR feat(dri3): serve DRI3 1.4 syncobjs through DRM #122); Phase A addresses the spec's genuinely-async flood case, which a conforming client can still hit.
2. Direct scanout requires the HW cursor (m1 guard cursor_hw). On nvidia-drm the HW cursor strategy is off by default, so Phase B does not engage on the nvidia box unless overridden (YSERVER_HW_CURSOR_NVIDIA=1, follow-up PR). This is the plan's known software-cursor blocker, recorded as a blocker rather than a regression.
3. Spec compliance goal with Xorg as ground truth — the async scrap mirrors present_vblank_scrap in Xorg's present.c.

Hardware validation (2026-08-12, nvidia box, CS2 no-vsync, Cinnamon)

• page_flip/s holds 59.8–61.0 through the flood (was 27–47).
• present_skips/s coalesces at a 266/s median.
• The stutter is gone; no visual or interaction regressions observed.
• m1_probe_pass=0 with no rejects/errors (HW cursor precondition unmet on this box).

Files

• crates/yserver-core/src/core_loop/process_request.rs — async defer/supersession + Skip eff=0.
• crates/yserver-core/src/present_scheduler.rs — successor/scrap scheduling.
• crates/yserver/src/kms/render/backend.rs — authoritative root, explicit-sync direct path, m1 pre-probe, composed-unflip degradation, late store ref.
• Docs: finding (2026-08-11-cs2-fullscreen-novsync-pageflip-collapse.md), transparency-bug finding, spec (2026-08-11-async-present-defer-supersession.md), plan (2026-08-11-fullscreen-direct-scanout.md), status.md.

Testing

• cargo test -p yserver --lib — 798 passed.
• cargo test -p yserver-core --lib — 1154 passed.
• cargo clippy --all-targets -- -D warnings — clean.

Dependency for reviewers

This branch is the base for the follow-up direct-scanout latest-wins supersession PR (feat/direct-scanout-latest-wins), which should be reviewed only after this one lands.