WIP: diag CI perf

Author: pimterry

.github/workflows/ci.yml

@@ -109,22 +109,58 @@ jobs:
           fi
         shell: bash
 
-      # [DIAG] arm64-only experiment to characterise the slow/variable Electron app-ready time and
-      # whether disabling the (failing) GPU init cuts & stabilises it. Runs regardless of the gating
-      # smoke test above (if: always()) and never fails the build (continue-on-error). Each iteration
-      # emits the full [DIAG][T+...] timeline incl. app-ready uptime. Remove with the rest of the diags.
-      - name: "[DIAG] arm64 startup-timing A/B (baseline vs --disable-gpu)"
-        if: ${{ always() && matrix.platform == 'linux' && matrix.arch == 'arm64' }}
+      # [DIAG] Cold-start breakdown: figure out *where* the slow first-launch time goes (Electron
+      # app-ready is ~14-20s cold vs ~0.7s warm). We deterministically reconstruct cold state with
+      # drop_caches + cache resets and isolate each suspected cost. Runs on both linux arches so we
+      # can also compare arm64 vs x64. Diagnostic only: always runs, never fails the build.
+      # Read the "[DIAG][T+... uptime=Xs] app ready" line emitted by each launch to compare conditions.
+      - name: "[DIAG] cold-start breakdown (fontconfig vs cold-reads vs first-run)"
+        if: ${{ always() && matrix.platform == 'linux' }}
         continue-on-error: true
         run: |
-          for i in 1 2 3; do
-            echo "===================== BASELINE run $i ====================="
-            xvfb-run --auto-servernum npm test || true
-          done
-          for i in 1 2 3; do
-            echo "================== DISABLE_GPU run $i ====================="
-            HTK_DIAG_DISABLE_GPU=true xvfb-run --auto-servernum npm test || true
-          done
+          set +e
+          ELECTRON_DIR="$PWD/node_modules/electron/dist"
+          USERDATA="$HOME/.config/HTTP Toolkit"
+
+          now_ms() { date +%s%3N; }
+          timed() { local s; s=$(now_ms); "$@"; local e; e=$(now_ms); echo "  -> $((e - s)) ms wall"; }
+          drop_caches() { sync; sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches' 2>/dev/null && echo "[page cache dropped]" || echo "[drop_caches FAILED - no sudo?]"; }
+          clear_fonts() { rm -rf "$HOME/.cache/fontconfig" 2>/dev/null; sudo rm -rf /var/cache/fontconfig/* 2>/dev/null; }
+          clear_userdata() { rm -rf "$USERDATA" 2>/dev/null; }
+          warm_electron() { find "$ELECTRON_DIR" -type f -print0 2>/dev/null | xargs -0 cat > /dev/null 2>&1; }
+          # Capture the full electron timeline ([DIAG][T+...]) + test markers ([DIAG][test...]),
+          # but drop the noisy renderer/CSP console spam. In cold runs 'app ready' is emitted before
+          # Playwright attaches its stdout handler, so grabbing the whole T+ timeline is more reliable.
+          launch() { xvfb-run --auto-servernum npx playwright test 2>&1 | grep -E "\[DIAG\]\[T\+|\[DIAG\]\[test" || true; }
+
+          echo "##################### HARDWARE / FILESYSTEM #####################"
+          uname -a; echo "nproc: $(nproc)"; free -h
+          grep -m1 -iE 'model name|CPU implementer|Features' /proc/cpuinfo
+          echo "--- workspace mount ---"; findmnt -T "$PWD" 2>/dev/null || df -h "$PWD"
+          echo "--- sizes ---"; du -sh "$ELECTRON_DIR" 2>/dev/null; du -sh "$PWD/node_modules" 2>/dev/null
+
+          echo "##################### RAW DISK READ of electron/dist #####################"
+          drop_caches
+          echo "cold read of electron/dist:"; timed warm_electron
+          echo "warm read of electron/dist:"; timed warm_electron
+
+          echo "##################### FONTCONFIG BUILD COST #####################"
+          clear_fonts
+          echo "fc-cache -f (cold build):"; timed fc-cache -f
+          echo "fc-cache -f (already built):"; timed fc-cache -f
+
+          echo "##################### ELECTRON app-ready BREAKDOWN #####################"
+          echo "Compare the 'uptime=Xs' at app-ready across conditions:"
+          echo "===== A: FULLY COLD (fonts cleared, userdata cleared, page cache dropped) ====="
+          clear_fonts; clear_userdata; drop_caches; launch
+          echo "===== B: fontconfig PRE-BUILT, else cold  (A minus B = fontconfig BUILD cost) ====="
+          clear_fonts; fc-cache -f >/dev/null 2>&1; clear_userdata; drop_caches; launch
+          echo "===== C: electron PRE-READ (page warm), fonts cleared  (A minus C = COLD-READ cost) ====="
+          clear_fonts; clear_userdata; warm_electron; launch
+          echo "===== E: fonts+page WARM, userdata cleared  (E minus D = Chromium FIRST-RUN cost) ====="
+          clear_userdata; warm_electron; launch
+          echo "===== D: FULLY WARM (everything cached) ====="
+          launch
         shell: bash
 
       - uses: actions/upload-artifact@v6