fix: bound readable content extraction

[chaliy]

Motivation

• The readable/agent content_focus path previously collected and cloned every nested content-like subtree, allowing a bounded but malicious HTML response to amplify into quadratic CPU and memory use and potentially OOM or stall workers.
• The change constrains work done during readability extraction to protect availability while preserving the goal of returning a dense article-like block for agents.

Description

• Replace whole-list candidate accumulation with single best-candidate tracking via collect_best_readable_candidate, scoring borrowed &str slices and cloning only the winning subtree to avoid retaining overlapping copies.
• Add a hard cap MAX_READABLE_CANDIDATES = 64 to stop scoring after a bounded number of candidates and stop scanning further low-priority candidates.
• Change score_readable_candidate to accept &str and return a lightweight ReadableCandidateScore, and only construct ReadableCandidate for the current best match.
• Add a regression unit test test_extract_readable_content_handles_deep_nested_content to ensure the extractor still selects readable content under deeply nested markup and preserve fallback behavior when confidence is low.

Testing

• Ran formatting with cargo fmt --all and static linting with cargo clippy --workspace --all-targets -- -D warnings, both completed without warnings or failures.
• Ran cargo test --workspace and all tests passed (300 passed; 0 failed), including the new regression test convert::tests::test_extract_readable_content_handles_deep_nested_content which passed.
• Verified the readable extraction behavior remains compatible with existing fallbacks (strip_boilerplate) when confidence is low.

---

Codex Task