KMunjy/maneno

Maneno — the first grapheme-aware crossword for Bantu languages (Shona, Swahili, Zulu). # bantu_crossword A grapheme-aware crossword generator for languages whose orthographies use **digraphs and trigraphs** as single units — built for Bantu languages (Shona, Swahili, Zulu, Xhosa, Tswana, Sotho) that don't have a strong crossword tradition and which standard "one letter = one cell" generators mangle. The core idea: a word is a sequence of **graphemes**, not characters. One grapheme fills one grid cell, and crossings happen at the grapheme level. So Shona `zvombo` is `zv · o · mb · o` (4 cells) and Swahili `ng'ombe` is `ng' · o · mb · e` (4 cells, apostrophe preserved). ## Scope **MVP1 targets Shona and Swahili** — the two easiest Bantu languages to generate crosswords for: regular orthographies, the smallest tokenizer edge-case surface (only Swahili's `ng'` apostrophe), vowel-heavy structure for dense crossings, and the best word-list availability. The Nguni group (Zulu/Xhosa) is deferred — clicks and trigraphs need tuning before they're puzzle-ready. `configs/zulu.json` is shipped as a **preview only** (flagged in its `_status` field), not an MVP1 target. ## Files | Path | Purpose | |------|---------| | `crossword.py` | Tokenizer + generator + renderer + CLI (stdlib only) | | `configs/shona.json`, `swahili.json` | MVP1 language rules | | `configs/zulu.json` | Deferred preview (not MVP1) | | `words/shona.json`, `words/swahili.json` | 10-word test datasets with clues | | `test_tokenizer.py` | Asserts digraphs occupy single cells | ## Usage ```bash python3 crossword.py --config configs/shona.json --words words/shona.json --show-solution python3 crossword.py --config configs/swahili.json --words words/swahili.json python3 test_tokenizer.py ``` ## Config structure ```json { "language": "Swahili", "atomic_graphemes": ["ng'", "ch", "sh", "gh", "ng", "ny", "mb", "nd"], "ignore_casing": true, "strip_diacritics": true, "preserve_marks": ["'"] } ``` - **`atomic_graphemes`** — sequences that occupy one cell. Order is irrelevant; the tokenizer always us …