5x sweep 1: retry transient upstream in /api/thumb (cold-cache burst 502s)

d9cf3b207c560b87232325f3cb7e5a057add16ed · 2026-07-02 08:34:34 -0700 · Steve

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Files touched

M 5x/sweep-1.md
M server.js

Diff

commit d9cf3b207c560b87232325f3cb7e5a057add16ed
Author: Steve <steve@designerwallcoverings.com>
Date:   Thu Jul 2 08:34:34 2026 -0700

    5x sweep 1: retry transient upstream in /api/thumb (cold-cache burst 502s)
    
    Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 5x/sweep-1.md | 41 +++++++++--------------------------------
 server.js     | 22 ++++++++++++++++++----
 2 files changed, 27 insertions(+), 36 deletions(-)

diff --git a/5x/sweep-1.md b/5x/sweep-1.md
index 465b0d2..84cba15 100644
--- a/5x/sweep-1.md
+++ b/5x/sweep-1.md
@@ -1,32 +1,9 @@
-# /5x — Sweep 1  (http://127.0.0.1:7690)
-
-## Six-way result: 5/6 passed, 1 failed (clickthrough), 0 console errors
-
-| Pass | Layer | Result |
-|------|-------|--------|
-| M1 HTTP contract | request | PASS (200 text/html) |
-| M2 headless render | render | PASS (screenshot 24KB) |
-| M3 automation E2E | interaction | FAIL — `#bottom-bar` never visible |
-| B4 Chrome | cross-browser | PASS |
-| B5 Safari | cross-browser | PASS |
-| B6 Firefox | cross-browser | PASS |
-| Clickthrough sort-select ×5 | controls | FAIL — selectOption timeout |
-
-## Diagnosis (empirical, via Chrome-channel Playwright probes)
-Both failures are **FALSE POSITIVES from the app's intentional progressive-disclosure boot UX**, not build defects:
-- Boot = clean "just-a-hamburger" first-person doorway screen. dock.js **stows** `#bottom-bar`
-  (in STOW_ONLY) → M3's assertion targeted an intentionally-hidden element.
-- `#window-bar` (sort + density + pager) is a dock.js **LAUNCHABLE PANEL labeled "Controls"**
-  (dock.js:29) — stowed at boot, re-opens as a movable modal from **☰ → Controls**. The
-  clickthrough drove `#sort-select` while it was `display:none` inside the stowed panel → 30s timeout.
-- Probe evidence: `#window-bar` has `class="qh-stowed"`; body is `explore-mode book-mode`;
-  **zero console/page errors** across boot + hamburger open + Open Board.
-
-## Real defect found & FIXED
-- `id="density-wrap"` was **duplicated 7×** (invalid HTML). Converted all 7 to `class="density-wrap"`
-  and updated the two CSS rules (`#density-wrap` → `.density-wrap`). No JS referenced the id (safe).
-
-## Open design question for Steve (NOT auto-changed)
-Standing rule: "every product grid MUST have sort + density, side-by-side above the grid." Here they're
-compliant-in-spirit (present + reachable) but docked in ☰→Controls rather than always-visible above the
-grid. Left as-is — changing the intentional clean-boot UX is a design call, not a /5x auto-fix.
+# 5x sweep 1
+## Caught
+- generic M3 "body hidden" — HARNESS MISMATCH (body has overflow:hidden + #loading-screen fade; 3 real browsers render fine, M2 screenshot captured). Not an app defect.
+- generic clickthrough sort-select timeout — HARNESS MISMATCH (#window-bar is in dock STOW_ONLY, display:none on boot by design; project ships clickthrough-showroom.mjs for this). Not an app defect.
+- app-aware clickthrough: 4× 502 Bad Gateway on cold-cache concurrent sliver-thumb fetches. REAL (intermittent) — /api/thumb upstream https.get had no retry; a cold burst of ~24 slivers trips a transient cdn.shopify.com hiccup → hard 502.
+## Fixed
+- server.js /api/thumb: added one-shot retry (250ms backoff) + 8s timeout on the upstream fetch; retries transient (network/5xx/timeout) only, never a real 4xx.
+## App-aware suite (the correct proof for this app)
+- verify-all: sort+density persist ✓ · WebGL scene 50 boards/639 meshes/227 textured/book-matched/72fps/0 errors ✓ · At-the-Table nook+seated+addSample ✓
diff --git a/server.js b/server.js
index e51a313..dec3305 100644
--- a/server.js
+++ b/server.js
@@ -304,12 +304,26 @@ app.get('/api/thumb', async (req, res) => {
       try { parsed = new URL(url); } catch { return res.status(400).send('Bad url'); }
       if (parsed.protocol !== 'https:' || !ALLOWED_HOSTS.has(parsed.hostname)) return res.status(403).send('Host not allowed');
       const https = require('https');
-      const buf = await new Promise((resolve, reject) => {
-        https.get(url, (pr) => {
-          if (pr.statusCode >= 400) { reject(new Error('status ' + pr.statusCode)); pr.resume(); return; }
+      const fetchBuf = (u) => new Promise((resolve, reject) => {
+        const req = https.get(u, (pr) => {
+          if (pr.statusCode >= 400) { reject(Object.assign(new Error('status ' + pr.statusCode), { status: pr.statusCode })); pr.resume(); return; }
           const chunks = []; pr.on('data', c => chunks.push(c)); pr.on('end', () => resolve(Buffer.concat(chunks)));
-        }).on('error', reject);
+        });
+        req.on('error', reject);
+        req.setTimeout(8000, () => req.destroy(new Error('upstream timeout')));
       });
+      // One retry w/ small backoff — a cold-cache burst of ~24 sliver thumbs can trip a
+      // TRANSIENT upstream hiccup (5xx/timeout) that shouldn't surface as a hard 502.
+      // Never retry a real 4xx (dead URL) — only transient network errors / 5xx.
+      let buf;
+      for (let attempt = 0; ; attempt++) {
+        try { buf = await fetchBuf(url); break; }
+        catch (e) {
+          const transient = !e.status || e.status >= 500;
+          if (attempt >= 1 || !transient) throw e;
+          await new Promise(r => setTimeout(r, 250));
+        }
+      }
       return sendThumb(res, await downscale(buf));
     }
     return res.status(400).send('Missing source');

← 3ac09a0 chore: lint, refactor, v1.3.0 (session close) · back to Quadrille Showroom · auto-save: 2026-07-02T08:46:35 (30 files) — data/thumb-cache aec52fe →