readability: extend score propagation for deeply nested SPA pages

[clementine-oaklight]

Problem

Modern SPA-rendered pages (React, Vue, etc.) wrap content in 3-5 levels of <div> wrappers. The readability scoring algorithm only propagated paragraph scores to parent (1×) and grandparent (½×), so the actual article container never accumulated enough score to be selected.

Result: On pages like Alibaba Cloud help docs, only a single paragraph (97 chars) was extracted instead of the full article (1118 chars).

Root cause

<div class="doc-body">                    ← 4 levels up, never scored
  <div class="rich-text-container">       ← 3 levels up, never scored
    <div class="section-wrapper">         ← grandparent, gets ½× score
      <div class="paragraph-wrapper">     ← parent, gets 1× score
        <p class="text-content">Content</p>  ← scored element

Each paragraph-wrapper gets scored independently and the best one wins — but it only contains one paragraph.

Fix

1. Extended score propagation from 2 to 4 ancestor levels with diminishing weights: [1.0, 0.5, 0.333, 0.25] (configurable via _ANCESTOR_WEIGHTS).

2. Lowered sibling inclusion threshold from best_score × 0.2 to best_score × 0.1, allowing more adjacent content sections to be included in the extracted article.

Test results

Page	Before	After
Synthetic Alibaba Cloud docs (deep nesting)	97 chars ❌	1118 chars ✅
GitHub Docs (cloning-a-repository)	4316 chars ✅	4316 chars ✅
Alibaba Cloud ECS instance families (7.1MB)	199821 chars ✅	199821 chars ✅

No regressions on traditional article pages.

[Oaklight]

Closing — superseded by #97 and subsequent changes on master. The deep nesting fix can be revisited separately if needed.