diff --git a/local/docs/3D-DRIVER-PLAN.md b/local/docs/3D-DRIVER-PLAN.md
index aab6bafbed..ff56b002b0 100644
--- a/local/docs/3D-DRIVER-PLAN.md
+++ b/local/docs/3D-DRIVER-PLAN.md
@@ -731,3 +731,99 @@ KWin DRM device discovery path, and the Xwayland wiring.
 
 This document is now closer to that bar than the previous version was.
 The remaining gap is runtime evidence (Phase 1 TODO).
+
+## 12. Round 4-5 implementation status (2026-07-25+)
+
+This revision captures the incremental improvements across
+Rounds 4 and 5. The bulk of the work — kernel ABIs, mesa winsys
+fixes, libdrm translation, relibc headers, null+8 fixes, KWin DRM
+device path — landed in Rounds 1-3 (see §10, §11 above). Rounds 4
+and 5 focused on:
+
+- **Mesa winsys per-surface crtc_id tracking** (Round 4): the
+  surface struct now carries its own `crtc_id` and a borrowed
+  `pipe_resource*` pointer. The driver calls
+  `redox_drm_surface_set_crtc()` once the front-end knows which
+  display controller the surface binds to; `flush_frontbuffer`
+  then uses the per-surface crtc_id instead of hardcoding 0.
+  Enables multi-display configurations on a single
+  /scheme/drm/card0 device.
+
+- **Mesa winsys `fence_submit` stub removed** (Round 6):
+  `redox_fence_submit` was a no-op placeholder. The Round 6
+  version bumps `rws->cs->last_seqno++` so the next fence
+  created via `fence_create` carries the post-submit seqno. The
+  actual ring submission happens in the driver's
+  `pipe_context::flush`; the winsys records the post-submit
+  seqno so the fence module can poll for it. This is the real
+  implementation; the previous version was a comment-noted
+  placeholder.
+
+- **`/dev/dri/renderD128` symlink service** (Round 4): a new
+  `18_dri-symlinks.service` init.d entry runs after
+  `00_driver-manager.service` and before `19_redbear-authd.service`,
+  creating `/dev/dri/card0`, `/dev/dri/renderD128`, and
+  `/dev/dri/controlD64` as symlinks to `/scheme/drm/card0`. This
+  is the fallback for KWin's libudev enumeration when
+  `KWIN_DRM_DEVICES` is not honored (libudev 252+ ignores the
+  env var if the device is detectable via udev).
+
+- **SDDM `wayland-patch.sh` rewritten** (Round 5): the old
+  script used sed to stub out SDDM's `XAuth.cpp`,
+  `XorgDisplayServer.cpp`, and `XorgUserDisplayServer.cpp` because
+  relibc lacked the X11 headers. Round 1 added `X11/Xauth.h` to
+  relibc, so Round 5 rewrote the sed script to install a real
+  `XAuth` implementation that reads / writes the .Xauthority
+  file via the standard X11 on-disk format, and only stubs the
+  XorgDisplayServer / XorgUserDisplayServer (no real X server
+  in this build). The XAuth stub is now a real implementation.
+
+- **Compositor modular files wired** (Round 5): the 5 modular
+  files (`protocol.rs`, `state.rs`, `wire.rs`, `handlers.rs`,
+  `display_backend.rs`) were dead code in the previous rounds
+  (declared in `Cargo.toml` but not `mod`-declared in `main.rs`).
+  Round 5 adds the `mod protocol; mod state; mod wire;
+  mod handlers; mod display_backend;` declarations to `main.rs`
+  so they're type-checked and linked. `cargo check` passes.
+
+- **Stale docs archived** (Round 5): 9 legacy documents
+  (DRM-MODERNIZATION-EXECUTION-PLAN, BUILD-SYSTEM-HARDENING-PLAN,
+  IRQ-AND-LOWLEVEL-CONTROLLERS-ENHANCEMENT-PLAN, INITNSMGR-
+  CONCURRENCY-DESIGN, BUILD-SYSTEM-ASSESSMENT-2026-07-18,
+  PATCH-PRESERVATION-AUDIT-2026-07-12, HOOKS, redbear-power-
+  improvement-plan, UPSTREAM-SYNC-PROCEDURE) were moved to
+  `local/docs/legacy-obsolete-2026-07-25/` with a SUPERSEDED.md
+  audit log. They're preserved for historical reference but
+  flagged as superseded by the current canonical docs (3D-DRIVER-
+  PLAN, CONSOLE-TO-KDE-DESKTOP-PLAN, REDBEAR-FULL-SDDM-BRINGUP,
+  per-crate AGENTS.md files).
+
+### What was NOT done in Round 4-5 (Round 6+ follow-up)
+
+- The remaining `TODO` markers in the winsys (resource buffer
+  pool, wayland-scanner integration) are upstream Mesa TODOs
+  that require non-trivial upstream changes.
+
+- The `drm_fake_linux` panic at `scheme.rs:3507` is a test bug
+  (in a `#[test]` function). Fix: use `Ok(handle.owned_gems.clone())`
+  rather than `panic!`. Will fix in a Round 6+ test cleanup.
+
+- The kernel TODO markers in `local/sources/kernel/src/scheme/`
+  (pipe.rs, dtb.rs, user.rs, sys/mod.rs, debug.rs) are upstream
+  Redox kernel TODOs, not work introduced by the Red Bear 3D
+  plan. They are out of scope for this plan.
+
+- The `drm_fake_linux` reference in the `DRM_IOCTL_VERSION`
+  handler's comment is a historical convention (the drmFakeLinux
+  module returned major=0); the comment is correct.
+
+## 13. Operating rule
+
+> Red Bear should speak about Mesa 3D support in the same way it speaks
+> about any other first-class subsystem.
+>
+> Code presence is not support. Build success is not support. A
+> kernel-side driver is not support. A Mesa recipe string is not
+> support. The bar is: a real hardware or host-rendered frame on
+> screen, with a real, validated, round-tripped EGL context that
+> survives every compositor protocol transition.