fix(translate): inline lead-in dup + dark-mode FAB + sidebar i18n (v3.5.38)

[heznpc]

Live debugging pass on anthropic.skilljar.com (extension loaded unpacked) surfaced five issues. All fixed; v3.5.37 → v3.5.38.

Fixed

1. Inline lead-in translation duplication (high impact) — a paragraph such as <strong>Estimated time:</strong> 15 minutes rendered as "Estimated time:예상 시간: 15분" (English original + translation side by side). safeReplaceText only rewrote direct-child text nodes, leaving text inside inline children (<strong>, <a>, <em>) untranslated next to the full-element translation. Any bold/linked lead-in hit this. It now writes the translation into the first meaningful descendant text node and blanks every other one — code/pre preserved, inline elements kept so links stay clickable. — src/content/content.js
2. Dark-mode FAB invisible — the AI Tutor launcher was #1a1a1a on a near-black page and disappeared; lightened accent + white icon + soft ring. — src/content/content.css
3. "Course roadmap" widget untranslated / un-themed — .crm-title / .crm-card-h were outside the translatable selector set; the .crm wrapper also stayed bright cream in dark mode. Both addressed. — src/content/content.js, src/content/content.css
4. Uncaught "Extension context invalidated" — resume.js touched chrome.storage.local unguarded, throwing after an extension reload/update while a tab was open. Now guarded + lastError-cleared. — src/content/resume.js
5. Stale sidebar labels after language switch — updateLocalizedLabels did not re-localize the Tools button, tools-menu items or example-question chips, so a post-build language switch froze them. Now re-applied (via DOM nodes, no innerHTML). — src/content/sidebar-chat.js

Tests / verification

• New tests/safe-replace-text.test.js — 5 cases incl. the exact bug repro, link preservation, code preservation.
• 488/488 jest pass; ESLint, Prettier, selector health, dict/i18n/glossary/bg-sync checks, docs-sync (idempotent), Firefox + bundle builds all clean locally.

Note (#1) was confirmed against the live DOM; the false "English-restore broken" lead from the debugging session was a measurement race and is not a real bug — restore works.