fix(playback): stop random black frames on software GL in frame-copy (#1748)

* test(e2e): explain black canvases in the packaged frame-copy smoke

When the smoke's canvas stays black, print and attach the session
snapshot (with mpv's drop counter), every slot of the helper's
shared-memory frame rings read from /dev/shm, and the session's verbose
mpv log (log-file). Together they separate these cases: the helper never
published a frame, mpv rendered black, or the preload pump did not draw
a real frame.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(playback): stop random black frames on software GL in frame-copy

mpv's LUT scalers (the default lanczos/spline family) build their weight
texture from an uninitialized talloc_array: mp_compute_lut fills only
kernel->size of every stride floats, so each row's padding is heap
garbage (reinit_scaler in video/out/gpu/video.c, mpv 0.41 and master).
The LUT is an rgba16f texture sampled with linear filtering. When the
padding holds NaN, or a value large enough to become Inf as half-float,
Mesa's llvmpipe carries it through the zero-weight neighbour texel, the
weights become NaN and the frame clamps to pure black. Whether it happens
depends on heap contents, so the packaged smoke's paused frame was black
in most attempts on the CI runner.

Evidence from CI (8 repeats each, no retries): baseline 2/8 passed;
LUT-free bilinear scalers 8/8; gpu-dumb-mode 8/8; LP_NUM_THREADS=1 6/8.
A synchronous glReadPixels right after mpv_render_context_render()
showed the black frame already in the helper's FBO. The readback, the
shared-memory ring and the preload pump were correct.

On a CPU rasterizer the helper now selects mpv's LUT-free bilinear
scalers and disables sigmoid upscaling. Software GL cannot afford the
LUT scalers anyway. A session option for the same key still wins.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(playback): report software scaler results and scope smoke ring diagnostics

The software-renderer fallback now reports which scaler options mpv
accepted and which it rejected (with mpv's error), instead of always
printing success. The smoke's black-canvas diagnostics read only the
failed session's rings (<sessionId>-g<N>), not every impv-fc ring in
/dev/shm.

Addresses Greptile review feedback.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: 4gray <fourgray@proton.me>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

docs/architecture/embedded-mpv-native.md

@@ -443,7 +443,12 @@ Trade-offs and constraints:
   string to stderr. If an early tier selects Mesa software rendering (for
   example, while a proprietary NVIDIA driver is reachable through the default
   display or GBM), it probes the remaining tiers and uses software only when
-  no hardware-backed context works. Windows (any arch with a helper, in
+  no hardware-backed context works. On a software renderer the helper sets
+  `scale`, `cscale` and `dscale` to `bilinear` and `sigmoid-upscaling=no`
+  unless a session option sets the same key: mpv's LUT scalers upload their
+  weight texture with uninitialized row padding (`reinit_scaler`, mpv 0.41),
+  and llvmpipe carries NaN/Inf from it through zero-weight filtering, so
+  sessions otherwise render pure black at random. Windows (any arch with a helper, in
   practice x64) is ported: WGL renders offscreen against a hidden window,
   the shm ring is a session-local named file mapping, and the reader addon
   compiles as C++ there (MSVC has no C11 `<stdatomic.h>`). The helper