Skin: per-pass GPU cost and memory for the skin stack — wall clock cannot resolve it #1396

Found while capturing the epic acceptance for #1222 (PR #1388).

The problem

#1222's delivery contract asks to "record image quality, per-pass cost and memory on named
hardware". PR #1388 delivers the image-quality half and a cost measurement that cannot resolve
the thing it measures.

DigitalHuman_Timing.txt times the full skin stack against a version-0 split-only control, per
raster path, wall clock, 384×384 offscreen, each arm after a discarded warm-up. Two consecutive
runs of the identical binary:

The full stack cannot really be cheaper than the control, so the negative values and the sign
flips mean the cost is below the fixture's noise floor — not that the stack is free. All that
file legitimately claims is that the stack does not cost an order of magnitude.

(An earlier revision reported a uniform +146%/+47%/+95%. That was the frame-graph reconfigure a
path switch forces, charged to whichever arm was captured first. Fixed in PR #1388; noted here so
nobody resurrects those numbers as a baseline.)

Memory was not measured at all.

What to build

A cost measurement that can actually resolve the skin stack:

• GPU timer queries, not wall clock. GPUPassTimerPool already exists and
  olo_render_frame_breakdown already reports per-pass GPU time; wall clock around RunFrames on
  a box that also hosts CI runners for another repo is the wrong instrument.
• A scene with enough skin to dominate the frame — one head at 384×384 is a rounding error.
  A crowd of heads, or a full-screen face at native resolution.
• Per-pass, not per-frame, so SkinDiffusionPass is attributable separately from the
  G-Buffer and lighting cost it rides with.
• Memory: the diffusion pass's transient targets and the profile table, on named hardware.

Acceptance criteria

• Per-pass GPU cost for the skin stack, with a stated run-to-run spread that is smaller than
  the effect being reported.
• Memory recorded for the skin-specific resources.
• Named hardware, and a re-run procedure someone else can follow.
• The result is a measurement to establish a budget from, per the epic's contract — not a
  justification for reducing ambition.

Evidence

#1222 (comment)

Score

capability: 3
craft: 5
stability: 3
decay: 2
effort: 5
confidence: 0.7
learning: 5
fun: 3
kano: performance
blocked_by: []
blocks: [1222]

_{Rated per issue-scoring · score = confidence × (capability + craft + stability + decay) / effort, derived by the picker.}