fix(eval): stop eval CI hanging past the timeout ceiling

.github/workflows/monthly-vision-eval.yml

@@ -12,18 +12,18 @@
 #
 # Why monthly not weekly: vision scoring costs real money
 # (Anthropic vision API, billed per image input token). At Haiku
-# defaults ~$0.001/image; 600 text images + 40 image-gen images
-# ~= $0.60-$1.00/run. Monthly cadence keeps the annual bill
+# defaults ~$0.001/image; 300 text images + 40 image-gen images
+# ~= $0.30-$0.40/run. Monthly cadence keeps the annual bill
 # bounded and still catches drift well ahead of any six-month
 # signal loss.
 #
-# Budget breakdown:
-#   - Text gen (600 samples):      free (CF OSS)
-#   - Text vision (600 images):    ~$0.60
+# Budget breakdown (5 models, n=30, raw + compiled = 300 samples):
+#   - Text gen (300 samples):      free (CF OSS)
+#   - Text vision (300 images):    ~$0.30
 #   - Image gen (40 images):       free (CF free tier, <10k
 #                                  neurons/day)
 #   - Image vision (40 images):    ~$0.04
-#   Total: ~$0.60-$1 / month
+#   Total: ~$0.30-$0.40 / month
 #
 # Secrets required on the repo:
 #   - CF_API_TOKEN + CF_ACCOUNT_ID: text generation via CF Workers AI
@@ -105,6 +105,7 @@ jobs:
             --brief briefs/landing.yml \
             --models "$MODELS" \
             --n "$N" \
+            --sample-concurrency 6 \
             --out evals \
             --report "docs/evals/monthly/${REPORT_DATE}-source.md"
           echo "REPORT_DATE=$REPORT_DATE" >> "$GITHUB_ENV"

.github/workflows/weekly-eval.yml

@@ -67,7 +67,7 @@ jobs:
           CF_API_TOKEN: ${{ secrets.CF_API_TOKEN }}
           CF_ACCOUNT_ID: ${{ secrets.CF_ACCOUNT_ID }}
         run: |
-          # Five OSS cells at n=30 = 300 raw + 300 compiled = 600 calls.
+          # Five OSS models at n=30 = 150 raw + 150 compiled = 300 calls.
           # Well within CF Workers AI free tier (10k neurons/day). If the
           # model list needs updating, bump the comma list below; the
           # workflow is intentionally explicit rather than auto-discovered
@@ -81,6 +81,7 @@ jobs:
             --brief briefs/landing.yml \
             --models cf:@cf/google/gemma-4-26b-a4b-it,cf:@cf/meta/llama-4-scout-17b-16e-instruct,cf:@cf/mistralai/mistral-small-3.1-24b-instruct,cf:@cf/openai/gpt-oss-120b,cf:@cf/qwen/qwen3-30b-a3b-fp8 \
             --n 30 \
+            --sample-concurrency 6 \
             --out "$OUT_DIR" \
             --report "$REPORT_PATH"
           echo "REPORT_PATH=$REPORT_PATH" >> "$GITHUB_ENV"

docs/artwork/stickers/README.md

@@ -0,0 +1,42 @@
+# Stickers
+
+Conference-distribution stickers for AHD. Two directions, both 3" × 3" die-cut matte vinyl, pure black on natural cream.
+
+## Files
+
+| File | Purpose |
+| --- | --- |
+| `02-rule-final.svg` | Editable source for *The Rule*. Live `<text>` elements, font references intact. |
+| `02-rule-print.svg` | Production. Text outlined to paths. No font dependency. |
+| `02-rule-print.pdf` | Production. Upload to printer. Vector, 3" page. |
+| `02-rule-print.png` | Preview only. 900 × 900 px (= 3" at 300 DPI). |
+| `03-manifest-final.svg` | Editable source for *The Manifest*. |
+| `03-manifest-print.svg` | Production. Text outlined. |
+| `03-manifest-print.pdf` | Production. Upload to printer. |
+| `03-manifest-print.png` | Preview only. |
+
+QR codes encode `https://ahd.adastra.computer` and are scanned-verified.
+
+## Regenerating from sources
+
+If you edit either `-final.svg`, regenerate the `-print` outputs with Inkscape under nix-shell so fonts resolve correctly:
+
+```sh
+nix-shell -E 'with import <nixpkgs> {}; let
+  fontsConf = makeFontsConf { fontDirectories = [ inter jetbrains-mono ]; };
+in mkShell {
+  buildInputs = [ inkscape fontconfig ];
+  shellHook = "export FONTCONFIG_FILE=" + fontsConf;
+}' --run '
+for stem in 02-rule-final 03-manifest-final; do
+  out="${stem%-final}-print"
+  inkscape "${stem}.svg" --export-text-to-path --export-plain-svg --export-type=svg --export-filename="${out}.svg"
+  inkscape "${out}.svg" --export-type=pdf --export-filename="${out}.pdf"
+  inkscape "${out}.svg" --export-type=png --export-dpi=300 --export-filename="${out}.png"
+done
+'
+```
+
+## Font note
+
+The editable sources reference Inter (SIL OFL) and JetBrains Mono (Apache 2.0). Both are free for commercial print. The live AHD site uses Neue Haas Grotesk (Linotype, paid); if you have a license, swap `font-family` in the `-final.svg` files before regenerating.

src/eval/runners/openai.ts

@@ -19,9 +19,20 @@ export function openaiRunner(options: {
    * fields to the underlying model as extra generation params.
    */
   extraBody?: Record<string, unknown>;
+  /**
+   * Per-request wall-clock cap in milliseconds. A request that hasn't
+   * resolved by this point is aborted so it surfaces as a caught error
+   * (the caller writes a `.error.txt` and moves on) rather than hanging
+   * the whole run. Without it a single stalled upstream connection
+   * blocks a serial eval forever — the failure mode that silently ate
+   * the full 60-minute CI ceiling. Defaults to 120s, comfortably above
+   * the slowest legitimate generation (~25-30s) observed on CF.
+   */
+  timeoutMs?: number;
 }): ModelRunner {
   const model = options.model ?? "gpt-5";
   const baseURL = options.baseURL ?? "https://api.openai.com/v1";
+  const timeoutMs = options.timeoutMs ?? 120_000;
   return {
     id: model,
     provider: "openai",
@@ -31,20 +42,29 @@ export function openaiRunner(options: {
       if (input.systemPrompt)
         messages.push({ role: "system", content: input.systemPrompt });
       messages.push({ role: "user", content: input.userPrompt });
-      const res = await fetch(`${baseURL}/chat/completions`, {
-        method: "POST",
-        headers: {
-          Authorization: `Bearer ${options.apiKey}`,
-          "content-type": "application/json",
-        },
-        body: JSON.stringify({
-          model,
-          messages,
-          max_completion_tokens: input.maxTokens ?? 4096,
-          seed: input.seed,
-          ...(options.extraBody ?? {}),
-        }),
-      });
+      let res: Response;
+      try {
+        res = await fetch(`${baseURL}/chat/completions`, {
+          method: "POST",
+          headers: {
+            Authorization: `Bearer ${options.apiKey}`,
+            "content-type": "application/json",
+          },
+          body: JSON.stringify({
+            model,
+            messages,
+            max_completion_tokens: input.maxTokens ?? 4096,
+            seed: input.seed,
+            ...(options.extraBody ?? {}),
+          }),
+          signal: AbortSignal.timeout(timeoutMs),
+        });
+      } catch (err) {
+        if (err instanceof Error && err.name === "TimeoutError") {
+          throw new Error(`openai ${model}: request timed out after ${timeoutMs}ms`);
+        }
+        throw err;
+      }
       if (!res.ok) {
         throw new Error(`openai ${model}: ${res.status} ${await res.text()}`);
       }

src/eval/runners/workers-ai.ts

@@ -8,7 +8,7 @@ export const WORKERS_AI_DEFAULTS = [
   "@cf/qwen/qwq-32b",
   "@cf/qwen/qwen2.5-coder-32b-instruct",
   "@cf/mistralai/mistral-small-3.1-24b-instruct",
-  "@cf/google/gemma-3-12b-it",
+  "@cf/google/gemma-4-26b-a4b-it",
 ] as const;
 
 // Model-family-specific generation knobs. Cloudflare Workers AI's