docs: sync ledger/blog with corrected numbers and the Arabic releases

[ronaldtse]

Summary

• /ml ledger: adds the two Arabic releases (ara-diac-1.0 server — 2.58 DER, best dedicated model on SadeedDiac-25; ara-diac-small-1.0 client — 8.26 full-set with the subset disclosure), corrects tha-g2p-base-1.0 to the runtime-protocol greedy numbers (3.53, teacher 1.25; the 9.19/4.43 beam-4 figures overstated error 4x), scopes the Persian claim to the SentenceBench homograph benchmark
• Blog: table rows corrected, title/intro bumped to eleven models, and a dated 2026-08-26 update section records the two Arabic releases, the protocol-matched leaderboard positioning (2.5793 vs GLM-5.2 2.6911 / Gemini-Flash 3.1926 / GPT-4 3.8645 / Sadeed 7.2915, behind only Claude 3.7 Sonnet's 1.3941), and the subset-lesson correction (client student 3.66 on 300 paragraphs → 8.26 on the full 1,200, teacher reproducing 2.5815 vs 2.5793 on the same run)
• All numbers trace to interscript-ml docs/RESULTS.md (PRs Improve Deployment Process #52/Improve Deployment Process #54)

Based on fix/pages-artifact-path (that branch carries the current site state); rebase to the migration target when it merges.

Test plan

• npm run check — 0 errors
• npx vitest run test/site.test.ts — 34 passed