benchmark: add live view CDP latency comparison tool and results

Benchmark tool and results comparing CDP operation latency across four
image variants: headless baseline, Approach 1 (Xvfb/noVNC, PR #174),
Approach 2 (CDP screencast, PR #176), and headful.

Covers 40+ CDP operations across 9 categories (screenshot, JS eval, DOM,
input, network, page, emulation, target, composite) plus concurrent load
testing. Includes results from Docker (4 vCPU / 1 GB headless, 8 vCPU /
8 GB headful), Docker with constrained headful (4 vCPU / 1 GB), and
KraftCloud (Unikraft) environments.

Key findings:
- Approach 2 (CDP screencast) adds near-zero overhead vs baseline
- Approach 1 (Xvfb/noVNC) adds ~30% overhead on input/screenshot ops
- Headful under headless constraints is not viable (39% idle memory)
- On Unikraft, Approach 2 remains the better choice for live view

Made-with: Cursor

Author: hiroTamada

benchmarks/liveview/.gitignore

@@ -0,0 +1 @@
+bench

benchmarks/liveview/README.md

@@ -0,0 +1,126 @@
+# Live View Benchmark
+
+Compares CDP operation latency and resource usage across four image variants:
+
+| Variant | Branch | Image | Description |
+|---|---|---|---|
+| `baseline` | `main` | headless | Current production headless (no live view) |
+| `approach1` | `feat/headless-live-view` | headless | Xvfb + noVNC live view (PR #174) |
+| `approach2` | `headless-cdp-live-view` | headless | CDP screencast live view (PR #176) |
+| `headful` | `main` | headful | Full headed browser (reference) |
+
+## Test sites (15 sites across 5 categories)
+
+| Category | Sites |
+|---|---|
+| Static | example.com, httpbin.org, lite.duckduckgo.com |
+| Content | Hacker News, Wikipedia, Reuters, Lobsters |
+| SPA | Google Maps, GitHub, Reddit |
+| Media | Unsplash, BBC News, CNN |
+| Complex | Amazon, eBay |
+
+## CDP operations tested (40+ operations across 9 categories)
+
+**Navigation** — full page load to each site category
+
+**Screenshots** — JPEG (q80), PNG, full-page capture, clipped region
+
+**JS Evaluation** — trivial expression, querySelectorAll count, innerText length, getComputedStyle, scroll to bottom, DOM manipulation (create/remove 100 elements), bounding rect collection
+
+**DOM** — getDocument at depth 1/5/full, querySelector, getOuterHTML
+
+**Input** — mouse move, click (press+release), type 21 chars, scroll (5x wheel events)
+
+**Network** — getCookies, fetch page body via Runtime.evaluate
+
+**Page** — reload (to load event), getNavigationHistory, getLayoutMetrics, printToPDF
+
+**Emulation** — set desktop viewport, toggle mobile viewport, set geolocation
+
+**Target** — getTargets, create+close tab round-trip
+
+**Composite scenarios** (realistic multi-step automations):
+- Navigate + screenshot
+- Scrape all links on page
+- Inject form + fill + read back
+- Click first link + wait for navigation
+- 10 rapid screenshots in sequence
+- Scroll 5 steps with screenshot at each position
+
+**Concurrent load test** — N parallel CDP connections all firing screenshots, evaluates, and DOM queries simultaneously for a configurable duration
+
+## Resource allocation
+
+Containers are constrained to match production allocations:
+
+| Type | Memory | CPUs | Variants |
+|---|---|---|---|
+| Headless | 1024 MB | 4 | baseline, approach1, approach2 |
+| Headful | 8192 MB | 8 | headful |
+
+Override with env vars:
+
+```bash
+HEADLESS_MEMORY=2048m HEADLESS_CPUS=4 HEADFUL_MEMORY=8192m HEADFUL_CPUS=8 ./run.sh
+```
+
+## Quick start
+
+```bash
+./run.sh
+
+# Customize
+ITERATIONS=5 WARMUP=2 CONCURRENT_WORKERS=5 CONCURRENT_DURATION=60s ./run.sh
+
+# Skip concurrent test for faster runs
+SKIP_CONCURRENT=true ITERATIONS=3 ./run.sh
+
+# Run specific variants only
+VARIANTS="baseline:main:headless: approach2:headless-cdp-live-view:headless:-e=ENABLE_LIVE_VIEW=true=-p=8080:8080" ./run.sh
+```
+
+## Requirements
+
+- Docker with BuildKit
+- Go 1.25+
+- python3 (for summary table generation; optional)
+
+## Output
+
+Results go to `results/<timestamp>/`:
+
+```
+results/20260310-143000/
+├── SUMMARY.md              # Combined comparison table
+├── baseline/
+│   ├── results.md          # Markdown latency table
+│   ├── results.json        # Machine-readable results
+│   ├── docker-stats.csv    # CPU/memory time series
+│   ├── memory.txt          # Idle + post-workload memory snapshots
+│   ├── container.log       # Full container logs
+│   └── bench.log           # Benchmark stderr
+├── approach1/
+├── approach2/
+└── headful/
+```
+
+## Running the Go benchmark standalone
+
+```bash
+cd server
+go build -o ../benchmarks/liveview/bench ../benchmarks/liveview/main.go
+
+# Against an already-running container
+../benchmarks/liveview/bench \
+  -cdp-url ws://127.0.0.1:9222 \
+  -iterations 3 \
+  -label mytest \
+  -concurrent-workers 3 \
+  -concurrent-duration 30s
+
+# JSON output
+../benchmarks/liveview/bench -cdp-url ws://127.0.0.1:9222 -iterations 3 -label mytest -json
+
+# Quick run without concurrent test
+../benchmarks/liveview/bench -cdp-url ws://127.0.0.1:9222 -iterations 2 -skip-concurrent -label quick
+```