fix(gpu,render): quiet the foreign-VRAM warning on the display card, pin pooled krea2/LTX-2.5 CLIP+VAE off it

[dmmdea]

Summary

Two operator-reported defects, one PR (clean-ship).

1. Foreign-GPU-memory warning too noisy (PR fix(gpu): strict FIFO lease queue, fair seat admission, clear-every-resident unload, foreign VRAM warning #469 follow-up). gpu status/gpu reserve's
   "foreign GPU memory holders" warning fired on ordinary desktop apps (Code.exe, PAIR.exe,
   chrome.exe, SnippingTool.exe, msedgewebview2.exe, explorer.exe, csrss.exe, opera.exe) sitting
   on the Qube's DISPLAY card (nvidia-smi index 1) — not a WORK card. Fixed:

   • per-process floor raised 64 -> 512 MiB, now configurable (foreign_gpu_min_mib)
   • OS shell/desktop denylist grows (csrss, explorer, shell-experience hosts)
   • the real fix: a resident whose every KNOWN GPU is a display card is dropped, but only on a
     box that also has a non-display card to score — a single-GPU box (its only card is also its
     display) still warns. Correlation is best-effort: PDH's per-process MiB (no per-card
     identity) cross-referenced against nvidia-smi compute-apps' per-pid gpu_uuid (reported even
     when used_memory is [N/A] on Windows/WDDM) and the existing gpuprobe.DisplayCardUUIDs.

2. Pooled krea2/LTX-2.5 text encoder loads onto the display card. Measured A/B
   (multigpu-archival-actions-2026-09-24.md): the stock CLIPLoader/VAELoader nodes'
   device input only offers "default"/"cpu" — "default" is ComfyUI's own fastest-first
   pick, independent of imagegen_pool_compute/imagegen_pool_donor, and on the 3x16 reference
   tier that IS the display card. Every pooled krea2 render put the ~12 GiB Qwen3-VL-4B text
   encoder there regardless of which two (non-display) cards the pool keys named. Fixed in
   render/wf-krea2.mjs and render/wf-ltx25-i2v.mjs: when pooled, the text encoder and VAE(s)
   now load through ComfyUI-MultiGPU's CLIPLoaderMultiGPU/VAELoaderMultiGPU, pinned to the
   pool's donor device (verified against the actually-installed plugin source on the Qube,
   C:\ComfyUI\custom_nodes\ComfyUI-MultiGPU\{__init__,wrappers}.py) — never the unnamed
   default. Unpooled builds are unchanged. render/comfy-nodes.mjs and
   internal/mediacap/routeneeds.go updated so doctor/MISSING_NODE name the two new classes.

How tested

• go build ./..., go test ./... — all green (root package, internal/config,
  internal/mediacap, everything else).
• node --test render/*.test.mjs — 443/443 green, including new pooled-CLIP/VAE-device
  assertions in wf-krea2.test.mjs/wf-ltx25-i2v.test.mjs.
• New Go tests: Qube-shaped (3 cards, desktop apps on the display card -> no warning; Resolve on
  a work card -> warning) and OptiPlex-shaped (1 card = display -> Resolve still warns) cases in
  gpu_foreign_test.go, plus effectiveForeignMinMiB coverage.
• Every new guard broken once (red) and restored, live in this session: the display-card filter
  in classifyForeign (removing it un-suppressed Code.exe on the Qube-shaped fixture) and the
  pooled-CLIP device pin in wf-krea2.mjs (reverting it re-failed both new JS assertions).
• config.example.json regenerated (go generate ./cmd/genexample) for the new
  foreign_gpu_min_mib key — caught by TestConfigExampleCarriesEveryKey, itself caught by the
  pr-review-toolkit:code-reviewer pass on this branch (its only finding; fixed).
• One pr-review-toolkit:code-reviewer (Sonnet) review of the whole branch: verified concurrency/
  timeout budget in gpu_foreign.go, the pid/GPU-UUID correlation edge cases, config wiring
  across all three gpu_cmd.go call sites, and consistency between the Go and JS node-pack
  tables. No other findings.

Risk

Low. gpu status/gpu reserve's warning is advisory-only (never gates a lease). The render
change is behind poolVvramGb > 0 (an explicit opt-in) and leaves unpooled builds byte-identical
— confirmed by tests. Neither path is exercised by CI hardware, so behavior is proven by unit
tests against the fixtures/measured values above, not a live render in CI.

🤖 Generated with Claude Code