PRD-570: a character costs what the camera sees of it

.github/workflows/native-platforms.yml

@@ -988,6 +988,38 @@ jobs:
         run: ccache --zero-stats
       - name: Build the native host
         run: pnpm --filter @threenative/runtime-native native:build
+      - name: Verify the shipped Crowd rate scenario on desktop
+        # Software adapters prove the native scenario contract, not physical GPU performance.
+        env:
+          TN_PLAYTEST_ALLOW_SOFTWARE: "1"
+          TN_FRAME_SPANS: "1"
+        shell: bash
+        run: |
+          set -euo pipefail
+          project="$GITHUB_WORKSPACE/packages/runtime-native/artifacts/conformance/crowd-rate-project"
+          mkdir -p "$project"
+          pnpm --filter skinned-crowd exec node ../../packages/runtime-native/scripts/bundle.mjs \
+            --project . --entry src/game.ts --target desktop --output "$project/skinned-crowd-native.js"
+          cp examples/skinned-crowd/public/assets.manifest.json "$project/assets.manifest.json"
+          timeout --kill-after=15s 315s node --input-type=module - \
+            "$GITHUB_WORKSPACE/examples/skinned-crowd/playtests/crowd-rate.playtest.json" \
+            --project "$project" --target desktop \
+            --executable "$GITHUB_WORKSPACE/packages/runtime-native/build/tn-linux/mystral" \
+            --host-arg run --host-arg "$project/skinned-crowd-native.js" \
+            --artifacts "$GITHUB_WORKSPACE/packages/runtime-native/artifacts/conformance/crowd-rate" <<'NODE'
+          import { main, flushStdout } from './packages/playtest/dist/runner/cli.js';
+          const keepalive = setInterval(() => {}, 1000);
+          const bound = setTimeout(() => process.exit(124), 300_000);
+          let code = 3;
+          try {
+            code = await main(process.argv.slice(2));
+          } finally {
+            clearTimeout(bound);
+            clearInterval(keepalive);
+          }
+          await flushStdout();
+          process.exit(code);
+          NODE
       - name: Report the compiler cache hit rate
         if: always() && env.CCACHE_UNAVAILABLE != '1'
         shell: bash

AGENTS.md

@@ -91,7 +91,7 @@ Results go in the existing PRD, PR or response; a separate evidence file only wh
 
 **Fail closed**: malformed input throws, a missing observation fails, an empty assertion set fails. `pnpm test` proves units; **a playtest scenario proves the game** — every runtime-behaviour change gets one, rerun on each later change to that behaviour.
 **Nearest lane first, CI last.** A CI round trip costs over an hour; prove it locally (unit test, playtest, emulator, attached device) and push once green. While `TN_RUNNER` is set, run only the focused checks for what you changed and push: the full board runs once, on `tn-local`. **Prefer an emulator over hardware** when it can hold the claim; a stopped emulator means start it, not "blocked". Never call `xvfb-run` (its exit status is its own cleanup kill; `sh scripts/xvfb.sh <cmd>` is the wrapper). An unnamed WebGPU adapter may be SwiftShader: pass `--browser-recipe webgpu` and check `adapter.info`. Free a port with `lsof -ti tcp:<port> | xargs -r kill`; `pkill -f vite` kills your own shell.
-**Visual judge subagent (owner rule, 2026-10-09).** Any change that alters what reaches the screen — materials, lights, shadows, exposure, fog, post, terrain, particles, LOD, resolution scale, and any perf change that touches them — needs a visual judge. Capture the same pose before and after (headed, `--browser-recipe webgpu`, `adapter.info` checked, at least two runs per arm, BEFORE and AFTER burned in large). Give both sets and the intended effect to a fresh read-only subagent that did not write the change. It returns IMPROVEMENT, NEUTRAL or REGRESSION per frame and overall. NEUTRAL or REGRESSION means fix the cause and judge again; do not explain it away. Record the verdict in the PRD box that claims the outcome and post the images on the PR (orphan `<prd>-screenshots` branch). A pixel diff or a green gate is not a verdict, and a perf win counts only when the judge passes.
+**Visual judge subagent (owner rule, 2026-10-09).** Any change that alters what reaches the screen — materials, lights, shadows, exposure, fog, post, terrain, particles, LOD, resolution scale, and any perf change that touches them — needs a visual judge. Capture the same pose before and after (headed, `--browser-recipe webgpu`, `adapter.info` checked, at least two runs per arm, BEFORE and AFTER burned in large). Give both sets and the intended effect to a fresh read-only subagent that did not write the change. Set visual acceptance criteria from the task before judging. The judge returns IMPROVEMENT, NEUTRAL or REGRESSION against those criteria per frame and overall. Performance-only and appearance-preserving tasks pass with NEUTRAL or IMPROVEMENT when there is no visible regression; tasks promising visual improvement require IMPROVEMENT. Intentional appearance changes must meet their stated visual target. REGRESSION, or NEUTRAL when improvement is required, means fix the cause and judge again; do not explain it away. Record the verdict in the PRD box that claims the outcome and post the images on the PR (orphan `<prd>-screenshots` branch). A pixel diff or a green gate is not a verdict, and a perf win counts only when the judge passes.
 Long gates write `artifacts/gates/status.json`: read it with `pnpm gate:status`, probe a stale phase with `pnpm gate:doctor`; `pnpm gate:resume` continues only while worktree, branch, HEAD, lease and artifact still match. None of them repair or remove a worktree.
 
 ## Working outside the repo

CLAUDE.md

@@ -93,7 +93,7 @@ Results go in the existing PRD, PR or response; a separate evidence file only wh
 
 **Fail closed**: malformed input throws, a missing observation fails, an empty assertion set fails. `pnpm test` proves units; **a playtest scenario proves the game** — every runtime-behaviour change gets one, rerun on each later change to that behaviour.
 **Nearest lane first, CI last.** A CI round trip costs over an hour; prove it locally (unit test, playtest, emulator, attached device) and push once green. While `TN_RUNNER` is set, run only the focused checks for what you changed and push: the full board runs once, on `tn-local`. **Prefer an emulator over hardware** when it can hold the claim; a stopped emulator means start it, not "blocked". Never call `xvfb-run` (its exit status is its own cleanup kill; `sh scripts/xvfb.sh <cmd>` is the wrapper). An unnamed WebGPU adapter may be SwiftShader: pass `--browser-recipe webgpu` and check `adapter.info`. Free a port with `lsof -ti tcp:<port> | xargs -r kill`; `pkill -f vite` kills your own shell.
-**Visual judge subagent (owner rule, 2026-10-09).** Any change that alters what reaches the screen — materials, lights, shadows, exposure, fog, post, terrain, particles, LOD, resolution scale, and any perf change that touches them — needs a visual judge. Capture the same pose before and after (headed, `--browser-recipe webgpu`, `adapter.info` checked, at least two runs per arm, BEFORE and AFTER burned in large). Give both sets and the intended effect to a fresh read-only subagent that did not write the change. It returns IMPROVEMENT, NEUTRAL or REGRESSION per frame and overall. NEUTRAL or REGRESSION means fix the cause and judge again; do not explain it away. Record the verdict in the PRD box that claims the outcome and post the images on the PR (orphan `<prd>-screenshots` branch). A pixel diff or a green gate is not a verdict, and a perf win counts only when the judge passes.
+**Visual judge subagent (owner rule, 2026-10-09).** Any change that alters what reaches the screen — materials, lights, shadows, exposure, fog, post, terrain, particles, LOD, resolution scale, and any perf change that touches them — needs a visual judge. Capture the same pose before and after (headed, `--browser-recipe webgpu`, `adapter.info` checked, at least two runs per arm, BEFORE and AFTER burned in large). Give both sets and the intended effect to a fresh read-only subagent that did not write the change. Set visual acceptance criteria from the task before judging. The judge returns IMPROVEMENT, NEUTRAL or REGRESSION against those criteria per frame and overall. Performance-only and appearance-preserving tasks pass with NEUTRAL or IMPROVEMENT when there is no visible regression; tasks promising visual improvement require IMPROVEMENT. Intentional appearance changes must meet their stated visual target. REGRESSION, or NEUTRAL when improvement is required, means fix the cause and judge again; do not explain it away. Record the verdict in the PRD box that claims the outcome and post the images on the PR (orphan `<prd>-screenshots` branch). A pixel diff or a green gate is not a verdict, and a perf win counts only when the judge passes.
 Long gates write `artifacts/gates/status.json`: read it with `pnpm gate:status`, probe a stale phase with `pnpm gate:doctor`; `pnpm gate:resume` continues only while worktree, branch, HEAD, lease and artifact still match. None of them repair or remove a worktree.
 
 ## Working outside the repo