Harden API + fix bugs (v3.1)

[willscott-v2]

Summary

Stacked on #5 (SI design system). Fixes the security issues and bugs from the full code review. No visual changes — review this one for behavior.

Base is feature/si-design-system, not main so the diff stays focused on logic. Merge #5 first, then this retargets to main automatically.

Security

• SSRF guards (api/chunk.js): http/https only; DNS-resolve the host and reject private/reserved ranges (127/10/172.16/192.168, 169.254 cloud metadata, 100.64 CGNAT, IPv6 loopback/ULA/link-local); every redirect hop re-validated (manual redirect, max 5); 5MB response cap; content-type allowlist. The endpoint was an open fetch proxy — anyone could scrape through our Vercel IPs or probe internal addresses.
• XSS: escapeHtml now escapes " and '. It's used inside title="${fragment}" where fragment comes from the scraped page's heading id — a page with <h2 id='x" onmouseover=...'> could inject a live handler. Verified the payload now stays trapped in the attribute value.
• Slack injection (api/feedback.js): escape &<> in message/email/pageUrl/user-agent — kills <!channel> pings and <url|label> link spoofing. pageUrl must be http(s), length-capped.
• Input validation: allowlist chunkSize/strategy/extract/format/tokenizer → 400 instead of a confusing 500; overlap bounded 0–200.
• Error leak: stop returning raw error.message to clients.

Bugs

• Overlap slider was permanently stuck on "Auto" — the input handler's condition made the manual branch unreachable, so users could never set overlap from the UI. Verified fixed: dragging to 25 now sends overlap: 25.
• Dead fetch timeout — timeout: 30000 is ignored by node-fetch v3. Replaced with native fetch + AbortSignal.timeout(25s), and dropped node-fetch entirely (engines already >=18).
• extract=defuddle was a lie — it silently fell back to cheerio. Now fails 422.

Housekeeping

• Removed scripts/build.js + vercel buildCommand. The build step mutated public/index.html in place (how the live GTM ID got committed). GTM-4G43 confirmed identical to the grader/ontologizer shared container; hardcoded.
• Deleted ~1,900 lines of unused root design-system files (superseded by the canonical theme; recoverable from git).
• README: MIT license (matched to package.json), searchinfluence/getchunks URLs, removed stale v1.2 section, honest "is my data secure" FAQ.

Testing

Local Vercel-style harness + Playwright:

• SSRF: localhost / 127.0.0.1 / 169.254 metadata / 10.x / ftp:// / file:// all rejected 400
• Live: example.com (cheerio fallback), wikipedia http→https redirect followed → 12 big / 29 small chunks, markdown format
• Validation: bad chunkSize/overlap/type → 400; defuddle-only on thin page → 422
• Slider fix and XSS containment verified in-browser

Deferred (not this PR)

• Vercel WAF rate-limit rule (dashboard config — needs to be applied there)
• Fixture/snapshot tests for split/merge/overlap
• Nested-content extraction fix for the cheerio fallback path

🤖 Generated with Claude Code