fix(ai): bound niri's NVIDIA free-buffer pool (#477)

* fix(ai): bound niri's NVIDIA free-buffer pool

Assisted-by: DeepSeek Harness 0.2.1-rc.2 (deepseek-flash)

* docs(ai): trim the niri VRAM workaround comment

Assisted-by: DeepSeek Harness 0.2.1-rc.2 (deepseek-flash)

WORKAROUNDS.md

@@ -51,6 +51,7 @@ A setting that works around a bug or a hardware quirk.
 | WA-031 | `iommu=pt` kernel parameter — `hosts/idols-aquamarine/hardware-amd-gpu.nix`                                                                                         | ROCm's system requirements warn that multi-GPU systems may need IOMMU passthrough to prevent application hangs, and this host has two Radeon AI PRO R9700 cards. Added before the first inference run; not yet observed on gfx1201.                                                                                                                                                                                                                                                                                                                                                                                                   | Drop once a ROCm release states that the default IOMMU mode is safe for multi-GPU gfx1201, or when the host runs a single GPU                                                                                                                           | 2026-10-13 | on a ROCm bump, or on the first hang report                                                         | active |
 | WA-032 | `amdgpu.runpm=0` kernel parameter — `hosts/idols-aquamarine/hardware-amd-gpu.nix`                                                                                   | this board's DSDT advertises ATPX power control (`\_SB.PC00.GFX0.ATPX`) but never defines the `PEG0.PEGP._OFF`/`_ON` symbols its body calls, so the first idle resume of the headless R9700 leaves it in `runtime_status = error` (`resume of IP block <smu> failed -62`) until reboot, and amdgpu discards the ACPI error. Root cause, kernel-source trail and the measured field comparison: `hosts/idols-aquamarine/ROCM.md`.                                                                                                                                                                                                      | Drop once a BIOS update defines those ACPI methods (upstream releases 1023/1066 exist for this board; a flash changes PCR7, so re-enroll TPM2), or a kernel release stops trusting ATPX when its power methods are missing                              | 2026-10-14 | on a BIOS update or kernel bump                                                                     | active |
 | WA-036 | catppuccin eval-time ports kept in the home closure (`home.extraDependencies`) — `home/base/core/theme.nix`                                                         | catppuccin/nix reads some ports at eval time (`lib.importTOML "${sources.starship}/<flavor>.toml"`) and inlines the result, so the built port, catppuccin's build/install hooks and the fetched port source stay unreachable: every gc deletes them, and the deploy after it rebuilds them (build logs: 2026-10-16). `home.extraDependencies` records the port in the generation, which is a gc root. Also read at eval time upstream: `bottom`, `element` (option `element-desktop`), `firefox`, `rio`, `palette` (options `skim`, `tty`), `spotify-player`.                                                                         | Drop once catppuccin/nix stops reading port files at eval time, or those ports come from a binary cache                                                                                                                                                 | 2026-10-16 | on a catppuccin bump                                                                                | active |
+| WA-039 | niri's NVIDIA free-buffer pool bounded (`GLVidHeapReuseRatio` 0) — `hosts/idols-ai/hardware-nvidia.nix`                                                             | the driver does not return freed vidmem buffers to its free pool, so niri's VRAM grows to ~1.1 GiB instead of the ~100 MiB its NVIDIA notes expect (measured 1137 MiB after 24 h uptime; [NVIDIA/egl-wayland#126](https://github.com/NVIDIA/egl-wayland/issues/126))                                                                                                                                                                                                                                                                                                                                                                  | Drop once a driver stops growing niri's VRAM without the profile, or niri ships its own default                                                                                                                                                         | 2026-10-19 | on an nvidia driver bump                                                                            | active |
 
 If WA-008 freezes recur, the fallback is to restore the parameters commented out in
 `hosts/idols-ai/hardware-configuration.nix` —

hosts/idols-ai/hardware-nvidia.nix

@@ -35,6 +35,33 @@
     powerManagement.enable = true;
   };
 
+  # Bound the driver's free-buffer pool for niri only: it never returns freed
+  # vidmem, so niri's VRAM grows past 1 GiB (WA-039). Applies at niri's next start.
+  # https://github.com/niri-wm/niri/wiki/Nvidia#high-vram-usage-fix
+  environment.etc."nvidia/nvidia-application-profiles-rc.d/50-limit-free-buffer-pool-in-wayland-compositors.json".text =
+    builtins.toJSON {
+      rules = [
+        {
+          pattern = {
+            feature = "procname";
+            matches = "niri";
+          };
+          profile = "Limit Free Buffer Pool On Wayland Compositors";
+        }
+      ];
+      profiles = [
+        {
+          name = "Limit Free Buffer Pool On Wayland Compositors";
+          settings = [
+            {
+              key = "GLVidHeapReuseRatio";
+              value = 0;
+            }
+          ];
+        }
+      ];
+    };
+
   hardware.nvidia-container-toolkit.enable = true;
   hardware.graphics = {
     enable = true;