Build 4.2.0 · 14 November 2026 · மூ/ழூ glyph fix · local-only TXT and DOCX conversion
Tamil document conversion
Large files run away from the interface in limited worker threads, with package checks, memory caps, progress and cancellation.
uh{
ஜ் 0x7B
ஹ் 0xC0
ஸ்ரீ 0x24
ஷ் 0x5A
ஔ 0x78 0x73
க்ஷ 0x42
மூ 0x5F
ழூ 0x3E
bare ங 0x2021
வி 0xC9
ரி 0xC7
Core repertoire: 309 directly verified canonical clusters; 4 font-confirmed unsupported clusters (ஙி, ஙீ, ஞி, ஞீ). Unsupported Tamil is never silently substituted, and Strict mode blocks mixed-encoding export.
| Source → Target | Pipeline |
|---|---|
| SunTommy → Unicode | sunDecode (longest-match) |
| Unicode → LT-TM-Annamalai | annEncode |
| SunTommy → LT-TM-Annamalai | sunDecode → Unicode → annEncode |
| LT-TM-Annamalai → Unicode | ambiguity-aware dynamic parse; complete parses beat greedy fragments, and genuine collisions are reported with alternatives |
| Auto-detect | Unicode Tamil is recognized automatically. Every non-Tamil plain-text input requires an explicit source choice; signature-matched decoders are scored only to rank a suggestion, so English and similar legacy layouts are never silently reinterpreted. |
| Word documents | logical glyph tokens can cross style and hyperlink run boundaries; output is assigned to the first contributing run, wrapper/relationship markup is preserved, and every join is reported. |
| 28 external encodings ↔ Unicode Bamini, TSCII, TAB, TAM, Anjal, Vanavil, Mylai, Shreelipi (+AVID), SoftView, Boomi, Dinakaran, Dinamani, Dinathanthy, Kavipriya, Murasoli, Nakkeeran, Pallavar, Indoweb, Koeln, Libi, Old Vikatan, Webulagam, TACE16, Indica, Anu, Roman, Diacritic | left-to-right longest-match tries write into a separate buffer, so decoded text is never reprocessed. Reverse output maps are governed separately. Unverified output encodings are disabled unless explicitly enabled. |
Actual LT-TM-Annamalai font SHA-256: E93DDF4A…742CF7B7 · no spelling correction · all ambiguity and unsupported Tamil reported · strict export on by default.
An admitted large file should not freeze the interface in a current Chromium browser when opened through Batch · DOCX. TXT and DOCX work is serialized in one background worker. The page enforces conservative working-memory budgets, blocks oversized packages before decompression, validates malformed archives, releases source buffers after each phase, and offers cancellation. One large DOCX part can still take time inside its worker, but the page remains responsive.
| Risk | Control in this build | Status |
|---|---|---|
| Main-pane conversion blocks the UI | Live conversion pauses above 50,000 characters; larger TXT files are routed to the worker-backed batch flow. | Closed |
| Several large documents exhaust memory | Every conversion is serialized in one worker. TXT uses a measured 128× admission multiplier; DOCX uses compressed, expanded, XML-part, convertible-character and estimated-peak budgets. Total selected/output bytes are capped. | Managed |
| ZIP bomb or malformed DOCX | Entry count, offsets, encryption, expanded size, single-part/XML size, materially dangerous compression ratio, output length and CRC are checked before or immediately after extraction. | Closed |
| Conversion cannot be stopped | Cancel active terminates running workers, clears queued work and leaves files ready for a clean retry. | Closed |
| Unbounded browser memory | Device-aware hard limits and enforced preflight/expanded-package estimates reject work outside the budget. Browser memory accounting varies, and true streaming XML conversion is not available in a standalone file. | Managed |
| Priority | Finding | Disposition |
|---|---|---|
| P0 | LT-TM-Annamalai மூ and ழூ codes were swapped, so every மூ printed as ழூ (e.g. மூன்று). | Closed Build 4.2.0: மூ = 0x5F, ழூ = 0x3E, checked by rendering ANNAMALI.TTF at large size and against a Word-printed proof. Documents converted with 4.1.x should be re-converted from their source, or have > replaced with _ in LT-TM-Annamalai text (no Tamil word in the checked book used ழூ). |
| P1 | A double-typed SunTommy pulli key printed as a stray “;”. | Closed SunTommy has no semicolon, so a pulli that cannot attach is removed and every place is listed in the report for review. |
| P2 | LT-TM-Annamalai draws “ ” as oversized marks next to ‘ ’, so typed ‘…” pairs looked mismatched in print. | Managed “‘ ’ quotes” (on by default) writes single marks in LT-TM-Annamalai output; Unicode output is unchanged. |
| P3 | Proofreaders read the joined க்ஷ glyph as a mistake. | Managed “க்ஷ split” (on by default) writes க் + ஷ. |
| P0 | Sequential replacement and greedy parsing corrupted external glyph sequences. | Closed Non-reprocessing governed tries plus ambiguity-aware parsing. Across 2,719,533 supported two-token checks, 2,716,803 were exact and all 2,730 non-exact parses were reported; zero were silent. |
| P0 | Greedy LT-TM parsing silently chose the wrong sequence. | Closed Dynamic parsing; all 97,969 attempted two-token combinations checked with no silent change. |
| P0 | Legacy plain text could be passed through or decoded under a similar font layout. | Closed Unicode Tamil remains automatic; every non-Tamil TXT input—including English—waits for an explicit source choice. |
| P0 | Active mapping overrides discarded ambiguity reports and DOCX source spans. | Closed Span-preserving override scanning now retains decoder tokens, offsets and all blocking issues; duplicate active rules are themselves blocking. Build 4.1.1 also ignores incomplete rules and reports them before export; blank replacements cannot delete text. |
| P0 | “Keep English” silently preserved glyph codes or converted real English. | Closed It is off by default. Enabling it for any legacy source always creates a blocking review gate, even when the heuristic preserved no span. |
| P0 | Unverified input could be confidently misread as a verified layout. | Closed Signature-matched unverified maps may improve the ranked suggestion, but no non-Tamil plain-text layout is committed by auto mode. |
| P0 | Large paste or main-thread fallback could freeze the page. | Closed Main-pane guard, background conversion and worker-required threshold. |
| P1 | LT-TM evidence was inferred rather than read from the actual font. | Closed 309 canonical clusters directly verified against ANNAMALI.TTF; four absent forms explicitly unsupported. |
| P1 | DOCX glyphs split by formatting or hyperlinks were missed. | Closed Logical token joining preserves wrappers and reports every cross-boundary join. |
| P1 | Legacy TXT byte output was always UTF-8. | Closed Explicit UTF-8, exact Windows-1252 and DOCX export modes. |
| P1 | Large or hostile DOCX packages had no resource validation. | Closed Device-aware admission checks plus ZIP structure, expansion and CRC validation. |
| P1 | Download-all compressed large outputs on the interface thread. | Closed ZIP creation now runs in the cancelable worker and has a separate conservative output budget. |
| P2 | Four LT-TM forms—ஙி, ஙீ, ஞி, ஞீ—do not exist faithfully in the audited font. | Managed Preserved and reported; Strict mode blocks mixed output. A different target font is required. |
| P2 | Twenty-four legitimate LT-TM byte sequences have more than one Tamil parse. | Managed Alternatives are reported; Strict mode requires a human decision. |
| P2 | Most third-party output fonts are not independently rendered and verified. | Managed Unverified targets stay disabled until the user explicitly opts in. |
| P2 | Custom Word font aliases may not match the built-in detector. | Managed Batch font inventory and per-font assignment are available. |
| P2 | DOCX XML is transformed as whole document parts rather than as a true stream. | Managed Worker isolation and expanded-size/part limits bound the risk. A native or server-side streaming transformer is the next step for files above these limits. |
| P2 | Very large issue lists are capped in the visible report and no signed conversion receipt is produced. | Managed Strict remains blocked, but a downloadable full JSON audit and input/output hash receipt remain recommended enhancements. |
| P2 | Browser compatibility has not been certified beyond current Chromium-based testing. | Managed Missing Worker or compression support fails safely; formal Firefox and Safari fixtures remain necessary. |
| P3 | No linguistic spelling or grammar correction. | By design This is an encoding converter; text meaning is never rewritten. |
.docx and text files. Binary .doc, password-protected/encrypted packages, ZIP64 packages, embedded OLE objects, charts, SmartArt text and external altChunk content are not converted.E93DDF4A…742CF7B7). The font binary is not embedded or redistributed.Reading input rules replace glyph codes with Unicode before decoding; writing output rules replace Unicode with glyph codes while encoding to that font. Type literal characters or hex (0x86, 0x0B95 0x0BCD). Fastest path: click any highlighted code in the unmapped report and fill in the replacement. Rules apply live — panes, batch files and reports all use them.