Release History

Changelog

A record of improvements, fixes, and new features — newest first.

0.6.1

2026-04-16

Tibetan syllable parser falsely flagged **འོ** (achung + o-vowel) on valid forms such as **བྱའོ**, **པོའོ**, and **མཐོའོ**. The **འི** path is now generalized: shared `ACHUNG_SUFFIX_VOWELS` (ི, ོ), one detect/parse pipeline, and completeness rules aligned so two vowel marks are allowed when the second sits on suffix **འ**.
Spelling rules page: **རུ་** locative example updated from **རྒྱལ་ཁབ་རུ་** to **བཀའ་རུ་**. A **བ** suffix takes **ཏུ་** (e.g. **རྒྱལ་ཁབ་ཏུ་** “to/in the kingdom”), not **རུ་**, which only follows **འ** or no suffix—consistent with the particle rules in the spellchecker.

2026-04-15

Docker Compose startup failure caused by `confidence_score` index referencing a non-existent column in `schema.sql` and `001_add_spelling_reference.sql`

`TextInput` Clear and Submit buttons now use the shared `Button` component
`spellcheck.tsx` `useEffect`-to-reset-state anti-pattern replaced with `checkedText`/`isStale` derived flag
`PDFUpload` internal `selectedFile` state removed; parent-owned state passed as prop
`ResultLink` (spellcheck.tsx) and `DownloadButton` (JobStatus.tsx) consolidated into shared `PDFDownloadLink`

2026-04-15

Upload size cap (50 MB) on PDF endpoints; returns HTTP 413 on oversized files
`max_length=100_000` on text spellcheck input to prevent event-loop blocking
Async page limit restored: PDFs ≤ 15 pages sync, larger PDFs routed to background queue

`TibetanSpellChecker` and `DictionaryService` now shared as module-level singletons — DB corpus loaded once at startup instead of per request
CPU-bound PDF processing steps (`extract_pdf`, `annotate_pdf`, `build_docx`, `check_text`) offloaded to thread pool via `asyncio.to_thread` to keep the event loop unblocked

Unused `checkHealth` function and `PDFResultURLs`, `SpellCheckRequest` interfaces from frontend API client
Unused `EmailStr` import from spellcheck schemas; `email-validator` dependency dropped
Unused `pyctcdecode` and `thin-plate-spline` dependencies
Stale archive docs: `REFACTORING_SUMMARY.md`, `MIGRATION_COMPLETE.md`, `SETUP_COMPLETE.md`, `RTF_QUICK_START.md`
Stale conversational artifact from `WORD_CORPUS_PLAN.md`

2026-04-14

Dictionary-based word lookup (Phase 2) — structurally valid syllables are now checked against a corpus of attested Tibetan words and flagged as `unknown_word` if absent, surfaced as yellow warnings distinct from red structural errors
`spelling_reference` database table with normalized lookup, source tracking, dialect and Sanskrit flags, and four indexes for fast access
`DictionaryService` — loads corpus at startup, extracts every valid syllable from multi-syllable entries, O(1) frozenset lookup; degrades silently when no database is configured
Corpus build scripts: `build_corpus.py` with scaffolded extractors for THDL, Rangjung Yeshe, and Monlam; `seed_corpus.py` with ~55 hand-verified words for pipeline testing
`corpus_hit` field on Phase 1 structural errors — flags when a flagged syllable is nonetheless attested in the corpus, useful for monitoring potential false positives over time
Structured JSON logging on every spell check (`spellcheck_result` tag) with aggregate counts for analysis via `grep | jq`
`GET /api/v1/corpus/stats` endpoint

2026-04-01

PDF spellcheck pipeline — upload a Tibetan PDF, get back an annotated DOCX with each structurally invalid syllable marked; supports both digital PDFs and scanned documents via OCR fallback
Async job support for PDF processing — upload returns a job ID; frontend polls until complete
BDRC ONNX OCR engine for scanned PDFs, replacing Tesseract; model files excluded from version control
Broken-CMap detection for digital PDFs — automatically falls back to OCR when the PDF's character map is corrupt or missing (ADR-015)
OCR/copy-paste parity tests — fixture-based regression tests verifying that the OCR and digital extraction paths produce equivalent spellcheck results
PDF spell check UI — upload form, job progress polling, email capture, results display
Shared `PageTitle` and `SectionHeading` typography components adopted across static pages

Particle validation tightened: ས, ར, ཞིག, and ཅིག marked as lenient to reduce false positives; ཏུ handling improved; locative suggestions refined
PDF annotation and DOCX export refactored to syllable granularity
Digital PDF extraction migrated from pdfplumber to fitz for better CMap support

2026-03-12

Particle spelling rules — context-aware validation of relational, agentive, locative, and indefinite article particles against the suffix of the preceding word, with suggestions for the correct form
Excepted words — ནའང་ and ལའང་ are now recognized as grammatically valid by convention and no longer flagged
Tibetan numeral syllables (U+0F20–U+0F33) are now silently skipped; numbers no longer produce false positives or break particle context
Tibetan punctuation and mark characters (yig mgo openers ༄ ༅, decorative shad variants ༑ ༒, gter tsheg ༔, etc.) are now silently skipped
Spelling Rules page — learner-facing reference documenting syllable anatomy, particle rules by category, and an explanation of each error type
Changelog page — release notes published on the frontend at /changelog

ཧཧིབ་ was not being flagged as invalid; ཧ (ha) cannot act as a prefix and is now correctly detected as an unparsed character
"Relational" used throughout in place of "genitive" for the bdag gi sgra particle category, matching standard English terminology for this community

Subscript validation re-enabled after confirming the rule data is correct and no false positives occur on attested Tibetan words

2026-02-09

Tibetan syllable spellchecker — validates the structural correctness of Tibetan syllables against classical spelling rules
Validates all five prefix consonants (ག་ ད་ བ་ མ་ ར་) against valid root consonant combinations
Validates superscripts (ར་ ལ་ ས་), subscripts (ྱ ྲ ླ ྭ), suffixes, and post-suffixes
Detects impossible syllable structures: subscripts appearing after vowels or suffixes, multiple vowel groups, unusual mark positions
Detects syllables that exceed the maximum valid length
Web interface for checking Tibetan text inline
Wylie transliteration shown alongside each flagged syllable for reference