Compatibility qualification — 0.3 beta

Initial baseline qualified on 2026-09-19 against words-mk270-1f2f0fb, using the exact upstream-tests-v1 profile. The reference program is WORDS 1.99.0 from the maintained mk270 lineage, not a reconstruction of historical 1.97FC.

npm run verify runs the offline build, frozen backend hash check and portable regression suite. The released 0.3.0 beta passed 36 groups covering the engine, adapters, reader, offline storage, input, documents, search state and native exception/capacity behavior. The 0.3.1 frontend release passes 43 groups: the original 36 plus compact notation, three dictionary-order checks and three stable-page/manifest checks. The ordering checks retain every candidate, form and explanation across the 3,953-input reference corpus, preserve equal-band order and keep nonordinal frequency categories distinct. The search-state fixtures preserve fresh native observations for failed attempts, affix fallback, distinct dictionary/affix spelling comparisons and transformed part-of-speech checks. The 0.3.1 website passes bounded Chromium/WebKit qualification for eight complete browser-to-Node JSON comparisons, explicit upgrades, old-tab coexistence, cold offline restart, interrupted/corrupt downloads, cache repair, stable help pages and touch/viewport behavior. Physical mobile-device installation and full classical-author corpora remain outside this qualification. The expanded corpus contains Aeneid vocabulary, all unique forms, available examples derived from the source inflection rules and explicit fallback probes. Repeated forms are deduplicated. This is broad feature coverage, not every possible stem, rule, affix combination or malformed input. Exact corpus hashes are in tests/compatibility/manifest.json.

The 0.3 beta designation recognizes preliminary engineering reliability after the complete comparisons below. The backend baseline is frozen; full classical-text corpus work remains deferred.

Comparison contract

The 21 raw Ada observations remain byte-for-byte evidence. The TypeScript comparison removes terminal prompts and whitespace around each frame, then trims trailing whitespace/CR on each output line. It retains internal blank lines, candidate multiplicity, order, morphology, flags, citations, meanings, explanatory text, unknown results and trimming markers. It never compares an unordered set. The CLI intentionally omits the startup banner and interactive command interpreter. The browser's selectable dictionary-frequency order is outside this raw-output contract. The library reader and original text/JSON retain their source order; the browser reorders complete entries only within uninterrupted interpretation groups. See browser ordering.

Expanded qualification

The 0.2 candidate includes full fixed-wordlist acceptance, which extends the frozen baseline and records newly exposed compatibility differences. The old finite baseline remains passing; its success is not universal equivalence. Browser macron adaptation is separate from the legacy profile.

The 0.3 storage/reader checkpoint replays all 4,144,877 frozen inputs. All 4,144,873 comparable normal outputs match, with zero regressions and no new exceptions. This resolves all 510 differences in the preceding source-audit checkpoint, including its two trimming-marker regressions, and all 109,001 output differences in the original 0.2 baseline. The earlier 240 input/option fixtures remain passing; another 180 cases across six option profiles and all 343 effective native addon fix/connection fields are now retained as regression observations.

The following local repair restores native line-atomic analysis and exceptional termination. All 222 input/option session cases match two native builds, compared with 68 before the repair. These include 105 native failures and controls for blank reads, 2,500-byte boundaries, same-line and preceding-line output, case, punctuation, compounds and Unicode. Six synthetic states injected at the native grouping caller verify 12/13 forms, 40/41 groups and 99/100 parse slots. The native grouping routine itself is unchanged in this probe. The final 100th nominal/verb record can fail because Ada evaluates the next array access even after the active range ends.

The two original native-error inputs remain errors in the source algorithm. They now return explicit legacy-error results with the original diagnostic body; CLI stdout, empty stderr and exit code 0 reproduce the native outer handler. These are classified separately from successful normal analyses. Empty input and ! remain limitations of the original wordlist console comparator. The 0.3 beta qualification does not establish universal exceptional-path equivalence. See the result contract.

Observed legacy behavior retained

The subsequent capacity review qualifies ten synthetic native caller states: dictionary candidates at 80/81, inflections at 249/250/251, reduced records at 250/252, and QU processing at 249/250/251. It preserves failing-write timing, the retained 81st candidate count, and the QU/PACK read beyond a full 250-slot array while searching for its null sentinel. All ten states match, compared with six before this repair. These are injected caller/table states in an isolated native copy; the original matching routines are unchanged.

All 23,808 nonempty distinct English index spellings, plus 192 POS/trim/input probes, match both native builds: 24,000/24,000, versus 23,963 before repair, with no regressions. Five observations intentionally reproduce the original 500-hit overflow; they are failure outcomes, not successful lookups. The other resolved observations cover native QV conversion and input with no ASCII word. Another 512 Latin sessions across 16 option profiles match both native builds, including combined settings and multiline sequences. Portable tests also reverse request order. These finite profiles do not qualify every option/input combination.

The complete 4,144,877-query Latin replay after these repairs retains all 4,144,873 normal matches and both original native failure outcomes, with no new differences or port failures. The two original comparator exclusions remain separate. This result is the frozen 0.3 beta backend baseline.

• Unique entries retain the original null part initialization (X), even when their parse quality is a noun, pronoun or verb. Citation artifacts remain.
• A precomposed macron splits ASCII lookup fragments differently from its NFD form; no implicit diacritic folding or i/j/u/v modernization is added.
• The order of the original linked inflection list affects which duplicate survives. Reading the text in the opposite order changes visible results.
• Explanatory rows affect fallback control: inserting a trick/syncope marker can suppress further affix searches. This is reproduced, including unknown results.
• Rejected attempts can leave dictionary candidates for the subsequent Do_Only_Fixes pass. That state and the shared reduced stem are request-local and retained at native pass boundaries, including the original stem-length reduction and empty QU/PACK search behavior. Syncope suppression is scoped to the native stages. Affix comparison folds u/v only; dictionary matching also folds i/j. Prefix checks use the part of speech after any suffix transformation.
• Shared parse buffers retain inactive slots after pruning or rollback; later passes can expose them. Mutating loops use the original Ada range bounds.
• Analysis completes for every word in a native input line before ordinary results for that line are printed. A failure discards that pending line, retains previously completed lines and terminates subsequent processing. Two blank native reads also end a session; prompt width after one blank read affects unknown-word alignment even when the prompt itself is omitted.
• Addon reading preserves the native comment buffer's retained tail. In the frozen data, this gives the ordinary adjective cumque tackon an unusable effective spelling. The portable reader reproduces that original defect without changing the source data or adding a lexical exception.
• PACK/enclitic duplicate groups, the adjective-to-adverb fallback, period-sensitive abbreviations and compound display behavior retain their original irregularities.
• Original data, meanings, frequency labels and heuristic suggestions are retained regardless of modern linguistic expectations. No corrected deviations are active.

Limits and next qualification work

The first-stage deliverables are operational. This is a beta release, not a claim of complete equivalence for every possible WORDS input or philological correctness. The qualified profile uses only the canonical general dictionary; local/special dictionaries, editing modes, interactive menus, persistent option files and historical platform-specific side effects are not implemented adapters.

Custom option combinations, unqualified original buffer/exception paths, additional cross-platform runtimes and broader affix cross-products remain release qualification work. Nondefault options are explicitly labeled custom, not silently presented as the frozen profile. Native file hash differences do not change the canonical portable data; see reference boundaries. Host memory exhaustion, storage-error retries, native file-system failures and platform-specific exception-information strings are not reproduced by the API. The CLI exposes the analysis body through its own documented arguments and stdin protocol; it is not a clone of native interactive menus or native command-line invocation modes.

Future linguistic changes must use a reviewed deviation with evidence and tests. Full classical-author corpus comparison and stress testing remain future work. The beta designation does not establish those claims. Release citation metadata is maintained in CITATION.cff.