fix: preserve literal HTML entities

[code-yeongyu]

Summary

Fixes a follow-up content-fidelity issue in the reader-mode normalizer. DOM and Turndown paths already decode HTML once, so normalization must not decode a second time and turn literal examples such as <custom-element> into active-looking tags.

Changes

• Keep entity decoding only in the raw regex fallback path, where HTML has not been parsed by DOM/Turndown.
• Add a regression covering markdown and text extraction for literal entity examples.

QA / Evidence

• RED: npx vitest --run test/webfetch.test.ts failed before the fix on literal entity preservation. Artifact: /Users/yeongyu/local-workspaces/senpi/.omo/evidence/webfetch-tistory-worktree-pr/external-entity-red.txt.
• GREEN: npx vitest --run test/webfetch.test.ts passed 16 tests and npm run check passed. Artifact: /Users/yeongyu/local-workspaces/senpi/.omo/evidence/webfetch-tistory-worktree-pr/external-entity-green.txt.

Risks

• The raw fallback still decodes once after tag stripping, preserving existing behavior for simple entities in non-DOM fallback scenarios.

Secret safety

No secret-bearing logs, tokens, auth headers, cookies, or credentials are included.

---

Summary by cubic

Fixes double-decoding of HTML entities in reader-mode normalization so literals like <custom-element> and AT&T stay literal in markdown and text outputs.

• Bug Fixes
  • Remove entity decoding from markdown/plain-text normalization; decode only in the raw fallback path.
  • Normalize whitespace separately in the fallback, then decode once after tag stripping.
  • Add regression test covering markdown and text to ensure one decode layer is preserved.

^{Written for commit 971ac6b. Summary will update on new commits.}