Harden Worker, site shell, gate suite, and Loopy skill

[trollbot2012]

Summary

Hardening pass over the Worker, site shell, gate suite, and the Loopy skill. Every change was built in its own branch and independently re-verified: full gate suite rerun, each new check mutated to prove it fails when the fix is reverted, and a written review before merge. Baseline gates were green before this work except one: loop-library/scripts/check.mjs failed on any Windows clone with core.autocrlf=true (CRLF vs \n assertions).

Worker tests: 52 → 72, passing on Node 22 and Node 24.

Worker

• Minimum completion time is now enforced server-side (form_elapsed_ms, 1200 ms suggestions / 800 ms weekly), mirroring the honeypot's fake-success response. Previously client-only, so a direct POST skipped it.
• Vote writes are rate limited per signed-in account through the existing limiter binding (30/min), the one write path that had no limit.
• Public rendered surfaces (homepage, detail pages, catalogs, feed, sitemap, llms.txt, public /api/loops) send public, max-age=60, stale-while-revalidate=600 instead of no-store. 404s, 5xx fallbacks, admin, session, vote and OAuth responses stay no-store. The homepage now copies only an allowlist of shell headers, so a shell Set-Cookie can never land in a cached response.
• Atom feed timestamps are UTC (Z) instead of a hardcoded -07:00.
• Real tests for LoopCatalog and VoteStore against SQLite via node:sqlite (both classes previously ran only behind in-memory mocks); tests for the 10/day and weekly 5/hour limits and the origin-suffix branch.
• Dead helpers removed; npm run check now node --checks every source file.

Site

• Server-rendered rows carry data-upvotes, so first paint is already in "Most popular" order instead of reshuffling after /api/votes returns.
• Vote and share controls are role="group" with per-loop labels; copy/show-more buttons name the loop. Twitter card meta on learn/ and agents/. The form guard fails closed (buttons disabled, status shown) when the API origin is missing.
• Inter and IBM Plex Mono are self-hosted (latin subsets, OFL licences included, font-display: swap); the CSS already named them but nothing loaded them.
• Cache-bust tokens bumped for both styles.css and script.js (the Worker-rendered pages were on a different, stale script.js token; a new gate keeps them aligned).

Gate suite and CI

• .gitattributes pins LF and check.mjs normalises CRLF on read, so the documented pre-commit checks run on Windows.
• check.mjs now mirrors every skills/loopy/references/* file against skills/loop-library/references/* by directory listing, gates audit.md, validates examples/loop.json with the real normalizeLoopDocument, checks .dev.vars.example points at the .com callback host, and asserts every src/*.js is in the check script.
• The CI git diff --check step ran on a clean tree after npm ci and inspected nothing; it is replaced with git diff --check $(git hash-object -t tree /dev/null) HEAD, pinned identically in ci.yml, AGENTS.md, README.md and check.mjs.

Loopy skill

• One terminal-state vocabulary across SKILL.md and the references (the run-receipt enum); "stagnated" removed, "exhausted" added to the preflight.
• publish.md gets the untrusted-content clause the other paths already had.
• Loop Doctor gains a Not ready verdict for gaps that cannot be repaired from scoped evidence.
• agents/openai.yaml advertises Find first instead of forcing the crafting interview.
• The design rules are stated once (SKILL.md "Design the feedback cycle"); the preflight and Loop Doctor point at it. Both copies of SKILL.md are kept in step and gated.

Docs

• .dev.vars.example pointed OAuth at the redirect-only .ai host; examples/loop.json carried a categoryLabel the renderer never shows; README publish checklist lacked llms.txt; the SEO audit presented a June snapshot as current; CHANGELOG backfilled and extended.

Not included

• No deploy and no catalog data change. Live verification of the cache headers needs a Worker deploy from main per AGENTS.md.
• The here.now proxy's handling of origin Cache-Control is undocumented; the headers are correct at the Worker boundary regardless.

Test plan

• node --check loop-library/site/script.js
• node loop-library/scripts/check.mjs (also on a fresh core.autocrlf=true clone)
• npm --prefix loop-library/worker run check on Node 24.18 and Node 22.23 (72/72)
• JSON manifests, git diff --check, and the new repo-wide whitespace check
• diff -r skills/loopy/references skills/loop-library/references empty

🤖 Generated with Claude Code