Custom shaders (shader engine 7B, first 7C slices) plus capture, ignore_alpha, focus and spring fixes

README.md

@@ -36,7 +36,7 @@ Underneath the water is a real tiling window manager: fast keyboard-driven layou
 - **It actually looks like something.** Ripples on open, wave transitions between workspaces, glass and shadow on your windows, floating windows that drift like they're sitting on water. Every bit of it is a config toggle, and none of it costs you performance you'll notice.
 - **Two ways to work.** Stick with classic numbered workspaces, or switch to Ocean: one endless zoomable canvas where every window lives in real 2D space instead of a workspace number. Switch between them live, no restart.
 - **Runs on what you already own.** Measured at ~60-70MB PSS at idle with the full water stack on, ~63MB with nine frosted-glass windows -- about half of Hyprland on the same hardware -- and it only climbs past that if you turn on more than the defaults.
-- **Genuinely yours.** Every ripple, wave, border, and shadow is a value in a config file, hot-reloaded on save. Nothing about how it looks is hardcoded.
+- **Genuinely yours.** Every ripple, wave, border, and shadow is a value in a config file, hot-reloaded on save. Nothing about how it looks is hardcoded. If that's not enough, write your own GLSL effect for a window and chain up to four passes of it (opt-in, see [Custom shaders](DOCUMENTATION.md#custom-shaders)).
 
 ## Try it
 

TECHNICAL_REPORT.md

@@ -32,21 +32,23 @@ Current release: **0.90.104**, second major pre-release. 1.0 is intentionally re
 | IPC: request/response plus an event-stream subscribe mode | Done |
 | `tidectl` CLI, including `doctor`/`report` for bug triage | Done |
 | Water/decoration render stack: impulse ripples, wave workspace transitions, tiled/floating water-glass and per-app liquid frost glass, analytical shadows, animated gradient borders, rounded clipping, configurable window animations, move/resize viscosity, connected-vessel BSP resizing, opt-in floating sway, automatic depth/buoyancy, ambient caustics | Done; tiled/liquid frost is release-GLES nested-verified in Classic/Ocean, standalone AMD remains, and the earlier stack is standalone-AMD-verified |
+| Custom shaders (`shaders { }`, `shader "<name>" { }`): opt-in user `.frag` effects over a window's backdrop, up to four chained stages with `save`/`get` and extra texture bindings, per-window and per-workspace assignment, bounded budgets | Done through Phase 7B; compile/draw covered by Mesa software-GLES tests, not yet run in a nested or native session |
 | Classic Depth Deck (tiled-window park/swap recall) | Done, standalone-AMD-verified |
 | Ocean spatial engine: reefs, per-output cameras, sink/dredge/surface depth, bookmarks, freeform window detach, smart tiling, live Classic↔Ocean migration | Done, standalone-AMD-verified |
 | Ocean compass (off-screen urgent/deep glow cues) and whole-world overview minimap | Done, nested-verified only |
 | Continuous swim navigation (trackpad swipe between workspaces) | Done, real-touchpad-verified |
 | Floating-window ocean physics (`float_physics { tier = light|full }`): disturbance-driven bob/drift, `full` adds mass/collisions/a traveling wave field | Done, nested-verified only |
 | Ocean currents: bounded render-only downstream drift for visible unfocused floating windows, with focus/drag pause | Done, release-GLES nested-verified on AMD |
 | Weighted buoyancy: per-app render-only mass/sink for Classic and Ocean floaters, with Ocean flow attenuation | Done, release-GLES nested-verified on AMD under Ocean; Classic shares the validated render path |
-| Nvidia support | Nested (EGL/GLES) verified on a real RTX 3060; standalone DRM backend not yet run natively on Nvidia |
+| Nvidia support | Standalone DRM backend verified on a real RTX 3060 (proprietary 595.99.02, two outputs at mixed scale, 2026-09-25); nested verified earlier |
 | AUR package | Not yet |
 
 ## Hardware verification
 
 - **AMD**: primary development and test hardware. The standalone `udev`/DRM backend, the full water/decoration render stack, swim's real-touchpad gesture path, and Ocean's core navigation (reefs, cameras, freeform drag) are all verified here.
 - **Latest standalone AMD health pass (2026-08-13)**: release 0.90.72 on Renoir completed full-output and cursor-overlay screencopy, reversible workspace transitions with ordered IPC events, invalid-action/batch/eval handling, and 100 one-shot plus 20 interrupted subscription connections without descriptor or PSS growth. Doctor remained all-pass with no TideWM panic/error or coredump. One-window PSS settled at 86.8 MiB after the reload/capture pass. The configured 24-fps caustics were the dominant continuous CPU/GPU cost; see `report.md` for the measured A/B and its whole-GPU caveat.
-- **Nvidia**: nested backend verified on a real RTX 3060 (proprietary driver): clean EGL/GLES context, correct rendering, no crashes. The standalone DRM backend and its Nvidia overlay-plane workaround are unverified on real Nvidia hardware.
+- **Latest standalone NVIDIA pass (2026-09-25)**: 0.90.106 (`13b104e`, `fiw/testing-fixes`) on the RTX 3060 above: doctor all-pass; full-output and cursor-overlay `grim`; unknown actions, an invalid mixed batch, eval syntax errors and a runaway eval loop (execution budget) all failed cleanly; oversized, garbage and truncated IPC requests left the compositor running; 100 one-shot queries plus 20 interrupted subscribers left PSS and descriptors unchanged. Idle ~1% of one core, 131 MiB TideWM VRAM; nine frost windows ~0-1% CPU and 139 MiB. PSS plateaus near 180 MiB after warm-up, of which ~60-95 MiB is `/dev/nvidia*` mappings; open/close cycles returned to the same value with a flat heap, so this is not a leak.
+- **Nvidia**: standalone `udev`/DRM backend verified on a real RTX 3060 (proprietary 595.99.02 with the open kernel modules, Gentoo, kernel 7.2.7), logged in from the display manager as a real session. GLES renders on the dGPU with the Radeon iGPU excluded through `gpu { render = vendor:nvidia; exclude = [vendor:amd] }`; two outputs run together (2560×1440@144 at 1.25×, 1920×1080@180 at 1.0×). The overlay-plane workaround branch runs for every output on this driver and scanout is correct on both; no overlay-specific test was done. Frost glass, custom shaders, screencopy, xwayland-satellite, portal and drag and drop worked. Measured cost is in `report.md`. The nested backend was verified earlier on the same card.
 - **Hybrid GPU**: selection/exclusion, render-node feedback, cross-device import, hot add/remove, and VT lifecycle compile and unit-test clean. The required end-to-end AMD+Nvidia PRIME/offload run is pending on the handoff laptop; do not call this hardware-verified until a non-primary game buffer displays correctly there at sustained frame rate.
 - **Intel**: untested so far.
 - **Still nested-only**: Ocean compass/overview, floating-window ocean physics (both tiers), `canvas_pan_button`, `modifier_pan_fingers`.
@@ -133,7 +135,6 @@ sudo cp share/xdg-desktop-portal/tidewm-portals.conf /usr/share/xdg-desktop-port
 - **Non-water motion presets**: smooth window move/resize and workspace motion that works with `water_effects = false`, exposed as Wave-selectable presets and tunable fields. Exact feel and defaults require maintainer approval before implementation.
 - **Feel-tuning** across viscosity, sway, depth timings, cascade's drag feel, floating-window ocean physics, and the transition/ripple presets. All ship with working defaults; the actual feel still gets refined against real use.
 - **Standalone hardware pass** for what's still nested-only: the Ocean compass/overview, and both floating-window ocean physics tiers.
-- **Nvidia native run**: the standalone DRM backend and its overlay-plane workaround still need a real TTY session on Nvidia.
 - **AUR package**: build from source for now.
 - **Design-pending identity features** (parking lot, needs a design conversation before pickup): tide contexts and workspace depth-moods. Currents, Cascade's pour/drain visual, and non-sorting weighted buoyancy are implemented with their approved contracts.
 

report.md

@@ -129,6 +129,12 @@ The finding text below is the original audit evidence. It is intentionally retai
 
 ### Validation state
 
+- Live standalone NVIDIA pass on installed release 0.90.106 (`13b104e`, `fiw/testing-fixes` on top of `5c63b90`; RTX 3060, proprietary 595.99.02 with the open kernel modules, Gentoo, kernel 7.2.7; udev, logged in from the display manager; HDMI-A-2 2560×1440@144 at 1.25× and DP-4 1920×1080@180 at 1.0× as reported by IPC; `gpu { render = vendor:nvidia; exclude = [vendor:amd] }` on a machine that also has a Radeon iGPU). This is the first standalone DRM run on Nvidia. `doctor` was all-pass. Full-output and cursor-overlay `grim` captures succeeded per output and for the combined 4960×1440 desktop. Unknown actions, an invalid mixed batch (nothing executed), an eval syntax error and a runaway `while true` eval (stopped by the execution budget) all failed cleanly; a 2 MB request was dropped at the size cap, random bytes got an `invalid request` error, and a truncated request got an error once EOF handling landed (`e4867b1`). 100 one-shot queries plus 20 interrupted subscribers left PSS unchanged (187,957→187,957 KiB) and descriptors flat (97→97). Measurements (20 untouched 1 Hz samples each; VRAM from `nvidia-smi -q -d PIDS`, GPU busy from `nvidia-smi`, which reads high at the 210 MHz idle clock and is not comparable to amdgpu `gpu_busy_percent`): idle with one window ~1% of one core, PSS 179.5 MiB, 131 MiB TideWM VRAM, 309 MiB card total; nine frost-glass Kitty windows plus one ~0-1% CPU, 139 MiB TideWM VRAM (card 653 MiB, the rest is client buffers as on AMD), 4.6 MiB of backdrop texture; back to idle 133 MiB. An earlier run on `5c63b90` with larger tiles measured 156→196 MiB TideWM VRAM for the same nine windows (15.3 MiB of backdrops).
+  Memory: PSS rises once to a plateau and stays there. `smaps` splits it into ~75-79 MiB heap and ~60-95 MiB of `/dev/nvidia*` mappings (driver-owned GPU buffers mapped into the process, which amdgpu does not show as PSS). Three cycles of opening and closing eight frost windows returned to exactly the same PSS with a flat heap, so this is a high-water mark, not a leak. The doctor's decorated-tier soft limit (280 MiB) was reached after about 4 hours of use on the older build; on Nvidia the tier limits should probably count the driver mappings separately.
+  Found and fixed on the same branch, each with regression tests: client drag and drop never worked (`WaylandDndGrabHandler` was an empty impl, so Smithay cancelled every source; `2727de9`); glass and other captures drew the wallpaper at `1/scale` size on fractional outputs (`3286393`); `output { }` blocks were ignored on reload (`8ebb259`); `spawn` could not pass arguments containing spaces and tidectl consumed flags after the command (`6517964`); `cursor_hide_after = 0` was rejected (`182df1b`); doctor mislabeled card nodes (`490df73`). Automatic attention depth became opt-in (`c593da7`).
+  Still open from this pass: user-reported tearing on the HDMI output only, unchanged at scale 1.0, not reproduced as a TideWM fault (cable or HDMI bandwidth at 1440p144 suspected); PDF text selection in Firefox reported as improved but still imprecise; layer glass and layer shaders were not exercised (no layer-shell client running yet).
+  Screencast: portal screen sharing from native Vesktop (vesktop-bin 1.6.7, Electron on the Wayland ozone backend) through xdg-desktop-portal to the TideWM ScreenCast backend and PipeWire 1.6.8 worked on this Nvidia setup, confirmed visually by the maintainer; the portal reported `AvailableSourceTypes = 7`. A successful session leaves no log line: `src/screencast` only logs failures, so a share is invisible in the session log unless it goes wrong.
+
 - Live standalone AMD pass on installed release 0.90.101 (`384fe98`, Renoir, udev; all geometry and cadence read from IPC): `doctor --json` was all-pass with the expected TideWM portal, PipeWire, xwayland-satellite, and AMD render node. Full-output `grim` captures verified the normal desktop, an in-flight water transition, five-window BSP/master/cascade layouts, pseudo-tiling, floating conversion, invalid-config diagnostics/recovery, and an XWayland GLX client. Twenty rapid workspace reversals settled on workspace 1 with no transition texture retained; the original window remained tiled, focused, and non-fullscreen. Four disposable translucent Kitty clients mapped/unmapped cleanly; five-window frost/water measured 64.6 MiB PSS with 0.9 MiB of active backdrop texture. A 15-second static A/B disabling only 24 fps caustics and focused-border animation changed `animation_active` from true to false and compositor CPU from 8.30% to 1.55%; the three touched Wave files were restored byte-for-byte. An invalid Wave edit kept the previous live config, showed the exact file/line in the first-party overlay, and cleared after restoration. Two identical rounds of 700 one-shot queries plus 100 interrupted subscribers converged to the same 91,006 KiB PSS high-water mark, zero swap, and exactly 93 descriptors, showing bounded first-use/cache growth rather than continued leakage. `glxgears` through xwayland-satellite mapped, rendered at 58.53/58.32 fps against the live 60 Hz mode, and unmapped cleanly. No new TideWM panic/error or coredump appeared. P-12's real-DRM wake path is reconfirmed, but this stripped release exposes neither a caustics-present counter nor a usable external perf/BPF/AMD busy counter, so exact presented-caustics cadence remains pending.
 
 - Live standalone AMD pass on installed release 0.90.72 (`dfc00b7`, Renoir, udev, 1920×1080@60 as reported by IPC; no geometry or mode was supplied by the test): full-output `grim` and `grim -c` captures succeeded; a reversible workspace 1→2→1 switch emitted the expected focus/workspace sequence and left the original window state intact; unknown actions, an invalid mixed batch, and invalid Wave eval all failed cleanly; 100 one-shot queries plus 20 interrupted subscribers left PSS flat at 88,840→88,838 KiB, swap at zero, and descriptors non-growing (91→90). Post-test doctor remained all-pass with no TideWM panic/error or coredump. The live log exposed two config-lowering warnings (`workspace_motion_delay = 0ms`, bracket-list ripple triggers) and `tidectl --socket … report` exposed a global-argument parsing bug; 0.90.101 fixes all three with exact regression tests and an end-to-end release-`tidectl` report run. The original live config values were restored after the performance A/B, and workspace 1 was restored after transition testing.

share/shaders/blur-h.frag

@@ -0,0 +1,11 @@
+// Horizontal half of a separable Gaussian blur. Pair it with blur-v.frag
+// as the next stage. Params: radius (float, texels between taps).
+vec4 tide_effect(vec2 uv) {
+    vec2 offset = vec2(u_texel.x * radius, 0.0);
+    vec4 sum = texture2D(tex, uv) * 0.2270270270;
+    sum += texture2D(tex, uv + offset * 1.3846153846) * 0.3162162162;
+    sum += texture2D(tex, uv - offset * 1.3846153846) * 0.3162162162;
+    sum += texture2D(tex, uv + offset * 3.2307692308) * 0.0702702703;
+    sum += texture2D(tex, uv - offset * 3.2307692308) * 0.0702702703;
+    return sum;
+}

share/shaders/blur-v.frag

@@ -0,0 +1,11 @@
+// Vertical half of a separable Gaussian blur, run after blur-h.frag.
+// Params: radius (float, texels between taps).
+vec4 tide_effect(vec2 uv) {
+    vec2 offset = vec2(0.0, u_texel.y * radius);
+    vec4 sum = texture2D(tex, uv) * 0.2270270270;
+    sum += texture2D(tex, uv + offset * 1.3846153846) * 0.3162162162;
+    sum += texture2D(tex, uv - offset * 1.3846153846) * 0.3162162162;
+    sum += texture2D(tex, uv + offset * 3.2307692308) * 0.0702702703;
+    sum += texture2D(tex, uv - offset * 3.2307692308) * 0.0702702703;
+    return sum;
+}

share/shaders/edge-mix.frag

@@ -0,0 +1,7 @@
+// Shows `original` (a textures binding) near the window's edges and the
+// stage input in the middle. Params: margin (float, fraction of the size).
+vec4 tide_effect(vec2 uv) {
+    vec2 edge = min(uv, 1.0 - uv);
+    float inner = smoothstep(0.0, margin, min(edge.x, edge.y));
+    return mix(texture2D(original, uv), texture2D(tex, uv), inner);
+}

share/shaders/tint.frag

@@ -0,0 +1,7 @@
+// Pulls the backdrop toward `tint_color` by `strength`.
+// Params: strength (float), tint_color (color).
+vec4 tide_effect(vec2 uv) {
+    vec4 color = texture2D(tex, uv);
+    vec3 tinted = mix(color.rgb, tint_color.rgb * color.a, strength);
+    return vec4(tinted, color.a);
+}