Phase 1 of iteration 1. Index: ohm_gl_fix. Origin: fourier Phase 5 review handover, 2026-04-30 — that page identified Mali EGLImage import cost as the binding constraint on the browser-side GL consumer; ohm_gl_fix is the campaign that addresses it directly rather than via browser-side carrier patches.

On bbb_1080p30_h264.mp4 (SHA-16 dcf8a7170fbd49bb), with mpv –hwdec=v4l2request –vo=gpu-next, over a 60 s steady-state window, reduce frame drops from the baseline 1039 / 1440 (72 %) to within the transient-startup floor measured on the working gst v4l2slh264dec → waylandsink control path on the same clip and window.

Equivalent framing: EGLImage-import cost ceases to be the binding constraint on per-frame budget on Mali-G52 / RK3566.

• “Within the transient-startup floor” instead of literal zero.

mpv's gpu-next path drops a handful of frames at preroll while

  format negotiation, decoder priming, and first GLES texture
  allocation settle. The ''gst → waylandsink'' control path drops
  0 / 1440 today, so the realistic floor is 0–N where N is the
  post-fix steady-state count after preroll. Locking the goal to
  the control-path floor keeps it falsifiable without being
  unreachably literal.
* **"Binding constraint" instead of a CPU% number.** A CPU% target
  presumes a particular ratio of import cost to composite cost. If
  a partial fix unmasks libplacebo composite or NV12→RGB shader
  fillrate as the new bottleneck, the goal as written still answers
  ("did import stop binding the budget?") rather than re-litigating
  the metric.
* Phase 3 will validate or invalidate this hypothesis. Loop edge
  3→1 stays open per ''feedback_dev_process.md''.

• No dropped safety checks. Question redundant per-frame validations

where buffer identity is already established at both ends; do not

  eliminate end-of-pipe sanity checks.
* Prefer pointer / fd passing over buffer copies wherever a copy is
  found. EGLImage caching keyed by dmabuf fd identity is the
  structural fix this implies; per-frame fresh-import is the
  structural waste.
* Patches and changes must be aligned to upstream in syntax and
  semantics. No upstream submissions (PRs/MRs/bug reports) without
  explicit task; flagging upstream-worthy work to Markus is welcomed
  with a why.
* Scope per ''~/src/ohm_gl_fix/README.md'': panfrost
  EGLImage-from-dmabuf for NV12 with Rockchip-relevant DRM modifiers
  + EGLImage caching across V4L2 buffer-pool reuse. Out: V4L2
  decoder work (fourier), browser carriers (fourier), compositor
  overlay-plane promotion, Vulkan video.

• Primary clip + config (locked): bbb_1080p30_h264.mp4

(SHA-16 dcf8a7170fbd49bb),

  ''mpv --hwdec=v4l2request --vo=gpu-next'', 60 s window.
* **Control path (locked):** ''gst-launch-1.0'' v4l2slh264dec →
  waylandsink, paced ''sync=true'', 60 s window — used to define
  the transient-startup floor.
* **Generalisation clip (Phase 6 if primary is met):** named in
  Phase 4. 1080p60 H.264 is the candidate; locking choice once
  Phase 2 + 3 expose any clip-class sensitivity in the import
  path.

• Campaign README: ~/src/ohm_gl_fix/README.md
• Origin freeze: fourier Phase 5 review handover
• Process spec: feedback_dev_process.md — the 8-phase loop
• Sibling project: fourier

Phase 1 entry locked 2026-04-30 after a goal-vs-constraint separation review with Markus. Phase 2 (situation analysis) is the next entry.