Files
iptvnator/spikes/mpv-frame-copy/PORTING.md
T
4grayandClaude Fable 5 26271fc076 feat(embedded-mpv): frame-copy rendering engine (experimental, macOS Apple Silicon) (#1169)
* spike(embedded-mpv): frame-copy pipeline prototype (helper + shm ring + Electron viewer)

Standalone macOS spike for the frame-copy unification direction from the
2026-07-10 analysis: a helper process renders mpv offscreen into a GL FBO,
reads frames back through an async PBO ring, and publishes BGRA frames into
a 3-slot POSIX shm seqlock ring; a minimal Electron viewer copies the newest
frame via a plain-C N-API addon and uploads it to a WebGL canvas per rAF.

First numbers on M1 Pro (see spike README): 4K60 HEVC hwdec sustained at
60 fps end to end, ~1.2 ms shm copy + ~3.5 ms texture upload, ~10 ms
produce-to-upload age, zero torn frames. Remaining gates: weak hardware,
long-run pacing, HDR, latency flash test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): add RESULTS.md measurement log with M1 Pro baseline

Structured per-machine table with repro commands so the pending Intel Mac
and Windows iGPU runs can be appended and compared one-to-one.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): pacing/judder instrumentation + 50/25 fps and HDR gate results

Viewer now measures inter-frame intervals on both clocks (present side and
producer side): stddev/p99/max, late-frame counters vs the producer's median
interval, and a cumulative LONGRUN summary every 30 s. The addon exposes the
producer timestamp (produceMs) for this.

Measured on M1 Pro: 50 fps and 25 fps cadences are clean (late frames only
at startup; residual jitter is 120 Hz rAF grid quantization, bounded by one
display tick), and 4K25 HDR10 PQ/BT.2020 is tonemapped to SDR by mpv before
readback at full rate with unchanged copy costs. RESULTS.md carries the
tables and HDR-clip repro commands.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): record 10-minute 4K60 HEVC long-run results

Zero dropped frames and zero torn reads after the first-minute warmup over
~8.5 minutes; steady-state late frames (~0.4%) track the 12 s test clip's
--loop restarts, not the copy pipeline.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): viewport-scaling measurement + integration design draft

Confirms the render-at-viewport-size claim (4K source in a 720p FBO costs
720p: 0.17 ms readback / 0.16 ms copy / 0.17 ms upload at 60 fps) and adds
DESIGN.md — the draft integration architecture: per-session helper process
linking bundled libmpv on all platforms (finally full-featured + Wayland-
agnostic Linux), JSON-over-stdio control evolving the Linux wid protocol,
unchanged EmbeddedMpvSession renderer contract, shm generations for resize,
packaging via the existing vendored-runtime tooling, rollout behind its own
flag with the docked path as default.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): auto-detect Homebrew prefix and Node headers for Intel Macs

BREW_PREFIX was hardcoded to /opt/homebrew (Apple Silicon) and NODE_INC to
one nvm version; both now resolve via brew --prefix and the PATH node's
execPath, so the pending Intel Mac run needs no Makefile edits. README gets
a fresh-machine checklist.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): self-contained measurement bundle for machines without Node/pnpm

make-bundle.sh assembles a tarball with the spike sources, vendored N-API
headers (Makefile prefers them when present, so no Node install is needed),
pre-generated 4K HEVC/HDR10 test clips, and an official Electron dist
download for the target arch. collect-results.sh builds and runs the full
RESULTS.md scenario suite automatically (plus an optional --long 10-minute
run) and writes one results-<host>-<date>.txt to send back. Target-machine
prerequisites shrink to Xcode CLT + brew mpv — built for the pending Intel
Mac baseline run.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): support MacPorts libmpv and legacy-macOS bundles

Makefile and collect-results.sh now detect libmpv in the Homebrew prefix or
MacPorts /opt/local (Homebrew is unsupported on legacy macOS like High
Sierra; 'sudo port install mpv +libmpv' provides libmpv there). make-bundle
takes ELECTRON_VERSION/BUNDLE_SUFFIX overrides — Electron 27+ needs macOS
10.15, so High Sierra bundles ship Electron 26.6.10 (LSMinimumSystemVersion
10.13).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* spike(embedded-mpv): scope macOS frame-copy engine to Apple Silicon only

Owner decision 2026-07-10: skip Intel Mac measurements and gate the future
frame-copy engine on arm64. Intel Macs able to run the app at all are a
shrinking 2015-2020 cohort and keep the docked/external/web player paths;
the macOS hardware gate closes with the M1 Pro numbers, and remaining
hardware risk moves to the Windows/Linux ports.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(embedded-mpv): frame-copy helper process and shm frame reader (native layer)

iptvnator_mpv_helper: one-process-per-session libmpv host that renders
offscreen at viewport size (headless CGL + async PBO ring, validated in
spikes/mpv-frame-copy), publishes BGRA frames into a seqlock shm ring with
resize generations, plays audio directly, and speaks a stdio protocol —
tab-separated commands in, JSON events out. The snapshot event mirrors
NativeEmbeddedMpvSessionSnapshot; status semantics (END_FILE reasons,
eof-reached with keep-open, pause gated on loaded path, fatal-only status
flips) are ported from embedded_mpv.mm.

embedded_mpv_frame_reader.node: plain-C N-API reader the preload script
uses to memcpy the newest complete frame into a V8 ArrayBuffer (Electron's
memory cage forbids zero-copy). Stub exports off macOS.

Both build as extra binding.gyp targets through build-embedded-mpv.js; the
helper gets the same libmpv dependency-path rewrite + ad-hoc re-sign as the
addon and is validated by the forbidden-link check.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(embedded-mpv): frame-copy engine wiring in main process and preload

EmbeddedMpvFrameCopyAdapter implements the NativeEmbeddedMpvAddon surface
over a per-session helper process (spawn, stdio protocol, snapshot cache,
graceful quit->SIGTERM->SIGKILL teardown), so EmbeddedMpvNativeService
reuses its polling/diff/power-blocker/recording logic unchanged. The
IPTVNATOR_ENABLE_EMBEDDED_MPV_FRAME_COPY flag (darwin/arm64 only) routes
getAddon() to the adapter and reports engine: 'frame-copy' in support.

The preload frame pump loads the shm reader addon, copies the newest frame
once per rAF into a reused buffer, and uploads it to WebGL2 on the
renderer's canvas — no frame data crosses the contextBridge; the bridge
only gains attachEmbeddedMpvFrameView/detachEmbeddedMpvFrameView. The
experiment flag relaxes the window sandbox for that native require;
contextIsolation and nodeIntegration:false stay on.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(embedded-mpv): frame-copy canvas mode in the player component + docs

EmbeddedMpvPlayerComponent renders <canvas data-embedded-mpv-frame> when
support reports engine 'frame-copy' and the session controller starts/stops
the preload frame pump around the session lifecycle. The bounds provider
skips HIDDEN_BOUNDS and the popover cutout for this engine — the canvas is
ordinary DOM, dialogs and popovers stack above it natively; bounds sync
still drives the helper's render size. Adapter unit tests cover spawn args,
snapshot caching, shm generations, protocol encoding, unexpected-exit
mapping, and dispose escalation. Architecture doc and CLAUDE.md describe
the engine, its flag, and the sandbox trade-off.

Verified end to end in the built app (M1 Pro): engine detection, helper
spawn, lavfi playback onto the canvas via CDP-injected smoke — including an
orientation fix (helper FLIP_Y already yields texture-order rows; the pump
shader must not flip uv again).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(embedded-mpv): close helper stdin on dispose + lifecycle logging

Live testing surfaced a stray idle helper that survived a session switch;
until the root cause is pinned down, dispose now also closes the child's
stdin (the helper exits on EOF) as a second kill path besides quit ->
SIGTERM -> SIGKILL, and spawn/dispose/exit are logged with the session id
so leaks are attributable.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(embedded-mpv): reap sessions when the renderer reloads or crashes

Root cause of the stray idle helper found during live testing: session
teardown lives in the renderer's Angular lifecycle, which never runs on a
renderer crash or hard reload — the main process kept the session (and its
frame-copy helper process / native mpv handle) alive until app shutdown.
EmbeddedMpvNativeService now watches the main window's webContents for
render-process-gone and did-navigate (full reloads only; in-app Angular
routing emits did-navigate-in-page) and disposes every session. Applies to
both engines. Verified live: location.reload() during frame-copy playback
logs 'Disposing 1 session(s): renderer reloaded' and the helper exits
cleanly. Regression test drives both events against the service.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(embedded-mpv): Settings toggle for the frame-copy engine

New 'Embedded MPV: frame-copy engine' checkbox in Settings > Playback,
shown only when the machine can run it (macOS arm64 with the helper binary
present — support now reports frameCopyAvailable). The choice persists to
the main-process config store because the engine relaxes the window sandbox
for the preload frame pump, which is fixed at window creation: main.ts
reads the store before creating the window and sets the engine env var; an
explicitly set env var (including '0') always wins, and the UI shows a
restart hint while the saved choice differs from the active engine.
Localized in all 18 locales.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(embedded-mpv): aspect-fit rendering in the frame-copy helper

The helper now observes dwidth/dheight and renders its FBO at the
aspect-fit size of the video inside the requested viewport, bumping a shm
generation on change — letterbox bars are never baked into frames (the VOD
watch shell's ~2:1 box no longer shows black side bars; the canvas
background is transparent so the sides show the app surface, while
fullscreen keeps its black backdrop). Frames also get smaller than the
viewport when aspects differ, trimming copy cost. Aspect override changes
refit automatically. Snapshots now carry videoWidth/videoHeight, and the
adapter forwards IPTVNATOR_EMBEDDED_MPV_AUDIO_DELAY to mpv's audio-delay
for lip-sync tuning until proper calibration lands.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* feat(embedded-mpv): require frame-copy artifacts in macOS package validation + docs

macOS packages that ship embedded_mpv.node must also ship the
iptvnator_mpv_helper binary and the embedded_mpv_frame_reader.node addon —
they come out of the same binding.gyp run, and a package missing them would
silently lose the frame-copy engine. Covered in the package-identity test.
Architecture doc and CLAUDE.md document the Settings toggle, aspect-fit
rendering, audio-delay passthrough, and the renderer-reload session reaping.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(electron): inline TS helpers so the sandboxed preload keeps working

The frame pump's async/await (target es2015 + importHelpers) made webpack
externalize tslib in main.preload.js. Sandboxed preloads can only require
Electron's built-in module whitelist, so the entire preload script failed
to load and window.electron disappeared for every run without the
frame-copy flag. importHelpers:false for electron-backend keeps the preload
bundle self-contained — and future async code in preload can no longer
silently reintroduce the breakage. Verified live: sandboxed run now has the
bridge, reports engine 'native' and frameCopyAvailable true.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(embedded-mpv): commit the frame-copy analysis handoff + source inventory

The 2026-07-10 analysis that led to this branch now lives next to the spike
(spikes/mpv-frame-copy/ANALYSIS.md), and the architecture doc's What To
Commit section lists the frame-copy engine sources.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(embedded-mpv): address review findings on the frame-copy engine

- Stale pump attach can no longer win over a newer session: attach/detach
  bump a shared epoch and async attach waits re-check it after every await,
  so an attach for a replaced session aborts instead of installing itself
  (greptile P1).
- A failed frame-view attach (no canvas, no WebGL2, reader missing) now
  disposes the session and surfaces the error UI instead of leaving audio
  playing behind a black canvas (codex P2).
- A stale frame-copy opt-in without the helper binary falls back to the
  native engine instead of reporting embedded MPV unsupported, and the
  Settings checkbox stays visible while a saved opt-in exists so it can
  always be cleared (codex P2). Regression test covers the fallback.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(packaging): make the darwin frame-copy packaging test host-agnostic

On non-macOS CI hosts validatePackagedEmbeddedMpv also reports that macOS
link validation needs a macOS host, so the success-path assertion now
checks only the frame-copy artifact requirement instead of expecting an
empty error list.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(embedded-mpv): Windows/Linux porting handoff for the frame-copy engine

Self-contained entry point for porting sessions on other machines: current
state and coordination constraints, per-OS task lists (Linux EGL first,
then Windows WGL + named shm — the decisive iGPU perf gate), the
hard-won gotchas from the macOS integration (preload/tslib sandbox
breakage, V8 memory cage, frame orientation, stale-attach epoch, dispose
escalation, node-gyp naming, snapshot protocol semantics), testing
recipes, and the suggested milestone order.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(embedded-mpv): branching and merge strategy in the porting handoff

Port work goes to stacked branches off the frame-copy branch (PR base =
frame-copy branch, sequential merges, stack depth one), never into the
frozen PR #1169 branch itself.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* docs(embedded-mpv): drop stale uncommitted note from porting handoff

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

* fix(embedded-mpv): harden frame-copy helper startup

---------

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-07-15 19:48:07 +02:00

12 KiB
Raw Blame History

Frame-copy engine — Windows/Linux porting handoff

Handoff for future Claude/dev sessions on Windows and Linux machines (they won't have the originating Mac's local session memory — this file is the transfer). Written 2026-07-11 by the macOS session that built the engine; fold into DESIGN.md once both ports land.

State as of this handoff

  • Branch: claude/embedded-mpv-frame-copy-6ddc36, PR: https://github.com/4gray/iptvnator/pull/1169 (CI fully green).
  • The engine works end-to-end on macOS Apple Silicon: Settings toggle → restart → helper process renders mpv offscreen → shm ring → preload pump → WebGL canvas. Verified live with real IPTV + Stalker VOD.
  • Scope decision: macOS = arm64 only (Intel Macs keep the native engine). Windows/Linux are NOT ported yet — this branch on those OSes behaves exactly like master (helper doesn't build there, engine can't activate, env flag falls back to native).
  • Coordination: PR #1169 credits larsemig's idea (#1154 comment 4932807350) and proposes the series merge plan: merge the shared-controls subset (#1148/#1149/#1152–54) rebased, supersede immersive (#1150/#1151) with this engine. Do not touch the controls layer until that lands.

What "porting" means

Only the helper (and a small reader-addon branch) is platform-specific. The stdio protocol, shm layout, TS adapter, main-process service, preload pump, and Angular UI are shared and already shipped.

apps/electron-backend/native/helper/
├── mpv_frame_helper.cpp     # portable: protocol, mpv session, snapshots
├── frame_helper_io.h        # portable: TSV-in/JSON-out, percent-encoding
├── frame_shm.h              # layout portable; POSIX shm calls are not
└── frame_helper_render.h    # macOS-ONLY: CGL headless GL + PBO + shm write
apps/electron-backend/native/src/embedded_mpv_frame_reader.c
                             # real impl under #ifdef __APPLE__, stub elsewhere

Porting = give frame_helper_render.h a WGL/EGL twin, give the shm create/open a Windows twin, flip the TS gates, extend packaging.

Branching & merge strategy (do this, not "commit to the current branch")

  • Never commit port work into claude/embedded-mpv-frame-copy-6ddc36 — that branch backs PR #1169, which is a frozen review/testing target (review fixes only).
  • Create a port branch off it: claude/frame-copy-linux-port (Windows later: branch off the Linux port branch if it reuses its portable clock/shm refactors — likely yes — otherwise off the frame-copy branch).
  • Open the port PR with base = the frame-copy branch, so the diff shows only the port. Merge order: #1169 → Linux PR → Windows PR. When #1169 merges and its branch is deleted, GitHub retargets the stacked PR to master automatically.
  • Keep the stack at most one unmerged level deep; if #1169 gains review commits, rebase the port branch onto it early and often.
  • Commit incrementally within the port branch; land each platform's measurement rows in RESULTS.md in the same PR as its port.
  • If #1169 has already merged by the time you read this: branch off master instead and ignore the retargeting notes.

Per-platform task lists

Linux (do first — much closer to done)

  1. Render backend: headless EGL (EGL_PLATFORM_SURFACELESS_MESA / eglGetPlatformDisplay(EGL_PLATFORM_SURFACELESS_MESA) with fallback to GBM) + the same FBO/PBO/readback code. GL entry points via eglGetProcAddress in mpv's get_proc_address.
  2. shm: POSIX shm_open works as-is. The ONLY blocker in shared code: clock_gettime_nsec_np(CLOCK_MONOTONIC_RAW) is macOS-only — replace with a portable now_ns() (clock_gettime(CLOCK_MONOTONIC, ...)) in frame_helper_render.h, frame_shm users, and the reader addon. Keep producer/consumer on the SAME clock.
  3. Reader addon: change #ifdef __APPLE__ to also cover __linux__ (code is already POSIX apart from the clock call).
  4. binding.gyp: helper target gets a linux branch — link system libmpv for dev first (-lmpv); bundled-libmpv runtime is a later packaging task. NOTE: the LINUX ADDON must still not link libmpv (that rule is for in-process only — the helper is a separate process, linking is the whole point and is legal there).
  5. TS gates: isFrameCopyEngineActive/isFrameCopyAvailable in embedded-mpv-native.service.ts (drop the darwin/arm64-only condition for linux), EmbeddedMpvFrameCopyAdapter.isSupported, Settings copy + i18n descriptions currently say "macOS Apple Silicon only".
  6. Big prize: no window embedding → native Wayland just works; later bundle libmpv → no system-mpv-on-PATH requirement → Flatpak/Snap. Update the Linux support matrix in docs/architecture/embedded-mpv-native.md when this lands.

Windows

  1. Render backend: WGL headless — create a hidden message-only window + dummy pixel format, wglCreateContextAttribsARB 3.2 core, then the same FBO/PBO path. MPV_RENDER_API_TYPE_SW is the fallback bring-up if WGL fights (CPU render, still proves the pipeline).
  2. shm: CreateFileMapping(INVALID_HANDLE_VALUE, ...) + MapViewOfFile behind the same FrameShmHeader layout. Name mapping: Local\\impv-... (session-local namespace). Reader addon gets the #ifdef _WIN32 twin. Atomics: std::atomic<uint64_t> fine with MSVC; the C reader can use InterlockedCompareExchange-free plain _Atomic-equivalent via C11 <stdatomic.h> (clang-cl) or volatile+ MemoryBarrier — simplest is compiling the reader as C++ on Windows.
  3. Process control: child.kill('SIGTERM') on Windows is TerminateProcess (no graceful signal) — the quit command + stdin-EOF paths (already implemented) are the graceful route; keep the kill as the hard fallback. stdio pipes work unchanged.
  4. binding.gyp: helper .exe target under OS=="win" linking the existing import lib (LIBMPV_IMPORT_LIB env — see build-embedded-mpv.js Windows path). DLL resolution: the helper exe sits next to lib/ with the mpv DLL — either copy the DLL beside the exe at build time or call SetDllDirectory/AddDllDirectory at startup. Watch the documented import-library-vs-DLL-basename gotcha (embedded-mpv-native.md).
  5. TS gates + audio: same switches as Linux. WASAPI audio comes from mpv directly — nothing to do.
  6. This is the open PERFORMANCE gate: mid-range iGPU laptop numbers decide go/no-go (RESULTS.md has the methodology + reference M1 numbers: 4K60 sustained, ~10 ms produce→upload, zero torn frames).

Both platforms — shared chores

  • validatePackagedEmbeddedMpv in tools/packaging/embedded-mpv-packaging.cjs currently requires frame-copy artifacts on darwin only — extend per platform when artifacts ship. Keep tests host-agnostic (CI runs them on a Linux runner; asserting an empty error list for a darwin dir fails there with "link validation must run on a macOS host" — already fixed once, don't regress).
  • getMainWindowScaleFactor (Electron screen) is cross-platform — no work needed; the helper receives device pixels.
  • Sandbox story: the flag relaxes the BrowserWindow sandbox for the preload reader require. Same trade-off applies on Win/Linux. Revisit-before- default-on candidates are in the architecture doc.

Hard-won gotchas (do not rediscover these)

  1. Preload + tslib: repo tsconfig has target: es2015; ANY construct that emits TS helpers in preload code (async/await, object spread in downlevel positions) with importHelpers: true makes webpack externalize tslib → the sandboxed preload dies with module not found: tslib → window.electron disappears app-wide. apps/electron-backend/tsconfig.app.json now sets importHelpers: false — NEVER revert it. Symptom to recognize: "Unable to load preload script" in renderer console.
  2. V8 memory cage: napi_create_external_arraybuffer over shm aborts in Electron. The reader MUST memcpy into a V8 buffer. Budgeted (~1.2 ms at 4K).
  3. Frame orientation: helper renders with MPV_RENDER_PARAM_FLIP_Y=1 and glReadPixels reads rows bottom-up → the shm buffer is already in texture order. The pump shader samples with UN-flipped uv. Adding a second flip shows upside-down video (bug already made and fixed once).
  4. BGRA fast path: readback as GL_BGRA/GL_UNSIGNED_INT_8_8_8_8_REV, upload as RGBA, swizzle .bgr in the fragment shader. On Windows check whether BGRA readback stays the fast path per driver; measure, don't assume.
  5. Aspect: mpv reports unset video-aspect-override as "-1.000000" → normalize to "no". The helper aspect-fits the FBO to dwidth/dheight inside the viewport (no baked letterbox bars) and bumps a shm generation (<base>-g<N>) on every size change; the pump re-attaches via the FRAME_SOURCE_CHANGED event.
  6. Stale attach race: attach/detach bump a shared epoch in the pump; every await re-checks it. Keep that invariant if touching the pump.
  7. Lifecycle: dispose escalation is quit-command → stdin.end() (helper exits on EOF) → SIGTERM(500 ms) → SIGKILL(2 s). The SERVICE also reaps all sessions on render-process-gone/did-navigate (renderer crash or hard reload never runs Angular teardown — without this, helpers leak). Watch ps | grep iptvnator_mpv_helper during any manual test session.
  8. Stale opt-in: isFrameCopyEngineActive() requires the helper binary on disk; missing helper = silent fallback to native, and the Settings checkbox stays visible while the saved value is true so it can always be cleared.
  9. node-gyp naming: module targets emit <target_name>.node (no product_name needed); the helper uses the "type": "none" default + per-OS "type": "executable" override trick in binding.gyp.
  10. snapshot protocol: helper's snapshot JSON mirrors NativeEmbeddedMpvSessionSnapshot verbatim (volume 0..1, nullable duration/track ids, videoWidth/videoHeight when known). Status semantics are ported from embedded_mpv.mm — END_FILE reason mapping, eof-reached ⇒ ended (keep-open), pause gated on loadedPath, only fatal/load errors flip status. Don't invent new mappings.

Testing recipes

  • Helper standalone (no Electron): (printf 'load\turl=av://lavfi:testsrc2=size=640x360:rate=30\n'; sleep 5; printf 'quit\n') | ./iptvnator_mpv_helper --shm-base /impv-t --width 1280 --height 720 → expect shm generations, snapshot events at 4 Hz, aspect-fit generation after video loads.
  • Reader probe (any Node ≥18): node -e "const r=require('.../embedded_mpv_frame_reader.node'); const i=r.open('/impv-t-g2'); ..." → latestSeq() advancing + pixel min/max spread.
  • In-app: IPTVNATOR_ENABLE_EMBEDDED_MPV_FRAME_COPY=1 pnpm run serve:backend:embedded-mpv or the Settings toggle (+restart). Second parallel instance for CDP testing: build, then run electron dist/apps/electron-backend/main.js --remote-debugging-port=9223 --user-data-dir=/tmp/x with ELECTRON_IS_DEV=0 for the file:// renderer (dist package.json has no main field — point at main.js explicitly; a separate user-data-dir avoids the Chromium profile singleton).
  • Perf gate: follow RESULTS.md methodology (STATS/LONGRUN lines, present-interval sd/p99/late counters). Reference: M1 Pro tables therein. The spike harness in this directory is macOS-only; for Windows/Linux measure through the real app + helper stderr or port collect-results.sh.

Suggested milestone order

  1. Linux helper bring-up (EGL + portable clock) → lavfi smoke → in-app behind flag → measure.
  2. Windows helper bring-up (WGL, named shm, reader twin) → same ladder → iGPU laptop numbers = the decisive open gate.
  3. Packaging: per-platform artifact validation + runtime staging.
  4. Only then: revisit Linux bundled-libmpv + Flatpak/Snap story.