Restore translucent+blurred Firefox look without the video-playback performance hit

[nethum529]

Investigated 2026-08-09 by a Fable agent against Hyprland 0.56.1 live options, current wiki source, and this machine's client state. Firefox look itself (moz-mac, user.js, vertical tabs) landed in PR #20; this issue is the compositor side.

Problem

The original look — opacity 0.90 0.93 on firefox + global blur (size 3, passes 2) — was reverted on 2026-08-01: video playing in a translucent+blurred Firefox measured 96% iGPU / 24% dGPU vs 47% / 0% opaque. Root cause: video damages a large region every frame, and a translucent+blurred window forces Hyprland to re-blur everything behind it per frame. Current state (in repo): windowrule = opacity 1.0 1.0 + no_blur on for firefox/discord/brave (hyprland.conf ~117-130, with the measurement comment block).

Verified facts (0.56.1)

• decoration:blur:new_optimizations defaults ON; ignore_opacity defaults ON; xray exists, off.
• Window rules support per-window xray on, no_blur, and 3-value opacity "active inactive fullscreen".
• hyprctl clients -j exposes per-window inhibitingIdle — Firefox sets an idle inhibitor during media playback → scriptable 'video playing' signal.
• amdgpu load readable at /sys/class/drm/card0/device/gpu_busy_percent (card0=amdgpu, card1=nvidia).

Ranked workarounds

1. Per-window xray (primary): windowrule = xray on, match:class firefox + restore opacity 0.90 0.93. Blur then samples only layers below tiled windows (effectively wallpaper) — a static, cacheable backdrop, so video self-damage stops forcing re-blurs. Identical look for tiled Firefox; only divergence: a FLOATING Firefox shows blurred wallpaper instead of the windows under it. Expected near-opaque cost; the wiki only documents the floating case, so the tiled-cache benefit is the hypothesis the measurement phase tests.
2. Fullscreen-opaque third value (free, compose with 1): opacity 0.90 0.93 1.0 — fullscreen video (worst case) renders opaque; one line, zero cost.
3. inhibitingIdle watcher (fallback): tiny user-service polling hyprctl clients -j (~1 s) or socket2-event-driven; on a firefox client flipping inhibitingIdle: true, tag it and apply opacity 1.0 override via a tag rule; restore on false. Exact look idle, exact opaque perf during ANY playback (windowed included). Costs a moving part; ~1 s translucent at video start; muted autoplay may not inhibit idle (untested).
4. Chrome-only translucency via userChrome (last resort): translucent toolbar/tabs/sidebar, opaque content. Savings not guaranteed (blur pass is per-window) and fragile across Firefox releases.

Rejected: smaller blur size/passes (look degrades, cost still damage-scaled); per-window blur size (doesn't exist on 0.56); global xray (alters every floating window; per-window rule strictly better).

Plan

• Phase 1: firefox only — delete the firefox opacity 1.0 1.0 and no_blur lines, add windowrule = xray on, match:class firefox + windowrule = opacity 0.90 0.93 1.0, match:class firefox. Discord/brave stay opaque.
• Phase 2: measure (protocol below). Accept if windowed-video iGPU ≤ ~55% (opaque baseline 47%) and dGPU hardware decode stays engaged. Reject if >70%.
• Phase 3 (only on reject): implement the inhibitingIdle watcher, keep xray.

Traps

• xray requires new_optimizations (default on) — do not disable it while 'optimizing'.
• ignore_opacity defaults ON — if blur looks missing after restoring opacity, check this before stacking hacks.
• Rule order: last match wins. DELETE the old firefox opacity 1.0 1.0 / no_blur lines rather than appending after them — appending before changes nothing.
• Three windowrule syntaxes exist in the wild (classic conf = this repo, lua = current wiki, windowrulev2 = old forums). Use the repo's match:class style; run hyprctl configerrors after every edit.
• hyprland.conf is stowed: change home/.config/hypr/hyprland.conf in the repo AND sync live (on the laptop they are separate real files).
• Rules belong in the main conf's OVERRIDES area — never in ~/.local/share/ambxst/ (generated, rewritten by axctl).
• Update the 2026-08-01 comment block (lines ~117-130) to reflect the new approach instead of contradicting it.
• The ffwarm special-workspace pre-warmed Firefox window inherits the same class rules; remember it when counting firefox clients in the phase-3 watcher.
• Phase 3 only: verify set_prop/tagwindow syntax on 0.56 via hyprctl help first; wiki examples are lua-flavored.

Verification protocol

Baseline on the current opaque config first, then after each phase, same monitor layout:

1. Same 1080p60 YouTube video, windowed, on the AMD-driven monitor, 120 s.
2. iGPU: for i in $(seq 60); do cat /sys/class/drm/card0/device/gpu_busy_percent; sleep 2; done | awk '{s+=$1} END {print s/NR"%"}'
3. dGPU: nvidia-smi dmon -s u -c 60 (sm + dec columns).
4. uptime load before/after; confirm hyprctl clients -j | jq '.[] | select(.class=="firefox") | .inhibitingIdle' flips true during playback (validates the phase-3 signal too).
5. Repeat fullscreen — must match the opaque baseline exactly (validates the third opacity value).

[nethum529]

GPU cost summary (explicit, so nobody has to reassemble it from the prose):

configuration	iGPU (video windowed)	dGPU	note
Original look: opacity 0.90/0.93 + blur, no xray	96%	24% SM, 0% NVDEC (hw decode dead)	measured 2026-08-01, load avg 11.8
Current opaque + no_blur	47%	0%, NVDEC engaged	measured 2026-08-01 baseline
Phase 1: xray + 0.90 0.93 1.0	expected ≈ baseline; accept ≤55%, reject >70%	NVDEC must stay engaged	hypothesis under test — wiki only documents the floating-window xray saving
Fullscreen video (third opacity value = 1.0)	must equal opaque baseline exactly	—	zero-cost path
Phase 3 watcher (opaque during playback)	= baseline during playback, look preserved when idle	—	fallback
Idle desktop, global blur size 3 passes 2	~4%	—	why blur stays on everywhere else

Mechanism of the cost: compositing scales with (damage rate × blur work). Video damages a large surface every frame; translucency forces the blur of everything behind it to be recomputed per damaged frame. xray decouples that by sampling only the static wallpaper; the fullscreen-opaque value and the watcher instead remove translucency exactly when damage is highest.

[nethum529]

Permanent-solution track (beyond the workarounds)

Two tiers of 'permanent':

Tier 1 — permanent config (likely sufficient). If Phase 2 accepts, the xray + 3-value-opacity pair IS a permanent fix for this machine: pure declarative config, no watcher, no moving parts, survives Hyprland upgrades as long as the two rules keep parsing. It only stops being enough if a floating translucent Firefox over other windows becomes a daily pattern (xray shows wallpaper through it, not the windows).

Tier 2 — root fix upstream in Hyprland. The actual defect-shaped fact: a translucent window's OWN damage forces re-blurring content BEHIND it, even though that background did not change. The blur input is invalidated by the wrong damage source. A root fix is damage-scoped blur caching: keep the blurred-backdrop texture per window/region and invalidate it only on damage underneath the translucent window, not on self-damage. Video playback then costs one cached-texture composite per frame — the look survives at near-opaque cost, no xray semantics compromise, benefits every translucent app (kitty included).

Concrete path for whoever picks this up:

1. Search hyprwm/Hyprland issues/PRs for prior art on blur caching / self-damage invalidation (terms: 'blur damage', 'blur cache', 'new_optimizations', 'xray tiled'). new_optimizations already preblurs a shared framebuffer once per frame — establish precisely when 0.56 marks it dirty (renderer blur pass in src/render/, the blur FB dirty flag) and whether self-damage of a translucent window is what invalidates it.
2. If the invalidation is genuinely coarser than necessary: prototype the scoped invalidation behind a config flag (e.g. decoration:blur:static_backdrop or per-window rule), measure with this issue's protocol.
3. Upstream it — small, measurable, benefits everyone; exactly the shape of patch the hyprwm maintainers take when it comes with numbers. Worst case a maintainer explains why invalidation must be coarse (occlusion/stencil interactions), which converts Tier 1 from 'workaround' to 'the correct answer' with a citation.

Ordering: do NOT block Phase 1 on this. Ship Tier 1 now; open the upstream investigation as a separate follow-up once Phase 2 numbers exist — the measured before/after is the strongest part of any upstream proposal.