Harden tier bucketing and PDF fallback text

[Moshbbab]

Motivation

• Make district-tier bucketing robust when quantile edges are non-unique or distributions are small to avoid qcut failures during feature engineering.
• Ensure PDF report generation does not fail on systems without Arabic TrueType fonts by providing safe Latin-1 fallbacks while preserving Arabic output when fonts are available.
• Provide a reliable end-to-end CSV → model → PDF flow for automated testing and demos.

Description

• Hardened district tiering by computing quantile bins with pd.qcut(..., duplicates='drop') and added an assign_tiers helper that adapts the number of labels to available bins and falls back to ranking when necessary; changes made in hemmah_pro_ivs_2025.py (location feature engineering).
• Reworked PDF rendering to detect whether Arabic fonts (Amiri) were added and to use a safe_latin1 fallback plus a render helper so text written with FPDF will not raise UnicodeEncodeError when Arabic fonts are missing.
• Kept bilingual (Arabic / English) fallback strings in report sections so generated reports remain readable when Arabic fonts are unavailable.
• Minor robustness and logging additions to surface progress during data cleaning, model training and report generation.

Testing

• Installed runtime dependencies and executed an end-to-end automated script that: generated a synthetic sample_real_estate.csv, loaded it with HemmahDataEngine, ran ivs_quality_check(), executed clean_and_engineer(), prepared modeling data, trained multiple models with HemmahMLEngine.train_multiple_models(), performed a predict() on a sample input, and generated a PDF via HemmahReportGenerator.generate_pdf(); this automated test completed successfully.
• The final automated run produced these key outputs: total_records: 200, unique_pct: 100.0, predicted_price_per_sqm: 576.2103499146549, confidence_interval: {"lower": 489.77879742745665, "upper": 662.641902401853}, model_used: "Random Forest", r2_score: 0.43099558871266896, and generated test_hemmah_report.pdf.
• During iteration there were failing runs (qcut duplicate-edge errors and a Latin-1 encoding error) which were fixed by the changes above; the final automated test passed without errors.

---

Codex Task