fix: follow ups and QoL improvements

[raisedadead]

Summary

Follow-ups from the #27 review, plus one pre-existing bug and some CI wiring.

formatHintOutput re-parsed HTML in order to escape it. A parser discards what it doesn't model before escaping can preserve it as text, and applies content models that don't fit a plain-text hint. One cause, three defects:

input	before	after
Your <img src="cat.jpg"> is missing alt.	Your &lt;img /&gt; … — attributes dropped	Your &lt;img src="cat.jpg"&gt; …
Wrap it in <code><textarea></code> and try.	<code>…&lt;/code&gt; and try.</code> — swallowed closing tag	<code>&lt;textarea&gt;</code> and try.
Comments look like <!-- this -->.	Comments look like . — comment deleted	Comments look like &lt;!-- this --&gt;.

Tags are now found in the raw output and everything between them is escaped. Detecting tags before escaping is what stops the renderer mistaking model-written <code> for markup it produced itself. Drops the sanitize-html dependency.

Fixes

• Prompt-frame strip loops until stable — one pass let <stu<student_code>dent_code> reassemble into <student_code>.
• challengeType guard restored — unrecognised values map to undefined again, as before Refactor/stabilize hint contract #27. The return type and docstring both already claimed this. No HTTP-visible change.
• Response limits declared — MAX_HINT_RESPONSE_CHARS is derived (≤5 chars per escaped code point), matching the measured worst case. Documents element-context-only insertion, since "/' aren't escaped. Fastify doesn't enforce response maxLength; this is for API consumers.
• test:manual no longer fails on a healthy dev server — its 401/403 cases can't pass outside production/staging yet gated the exit code, and spent two Groq calls doing it. The runner now detects whether auth is enforced and skips them when it isn't.
• Sentry flush on startup failure (src/index.ts) — pre-existing, not from Refactor/stabilize hint contract #27. Boot failures were logged at error then dropped before the transport drained.

CI and tooling

• pnpm run verify (check && typecheck && test) is now the single source of truth; the CI workflow calls it. Local and CI can no longer drift — this PR's first CI failure was exactly that drift, pnpm lint locally vs pnpm run check in CI, so prettier never ran.
• typecheck now covers scripts/ via tsconfig.check.json (noEmit); tsconfig.json stays the build config and dist/ is unchanged. src/**/__tests__ stays excluded — vitest's axios mocking and dynamic import() don't satisfy moduleResolution: nodenext, so including them means rewriting mocks, not fixing bugs.
• oxlint ignores .scratchpad/dist/coverage, matching .prettierignore. Local file count now matches CI's.

CLAUDE.md updated throughout: #27 moved auth to onRequest, inverting the documented hook order.

Verification

pnpm run verify — 104 tests (up from 88), lint and prettier clean, typecheck clean; pnpm run build still emits the Lua script. Every behaviour change written test-first. Verified live against a dev server: contract runner exits 0; four /hint calls across html/css/javascript returned balanced <code> and no active non-code markup.

The chore:/ci: commits are independent of the fix: ones and can be cherry-picked out if you'd rather keep this PR purely about the hint formatting.