Full-consonant default contracts — 2026-10-03

Publication follow-up (2026-10-03): 0.1.0-alpha.2 was subsequently published on npm and its registry artifact verified against the reviewed source. See the release record for current distribution status. This dated record retains the source-candidate checks made before publication.

The source now defaults to consonantMode: 'full'. An unmarked final fallback consonant renders full; adjacent consonants still form automatic clusters. createEngine({ consonantMode: 'half' }) retains the previous implicit final-halant behavior. Slash and backtick explicitly request a halant, and /= and backtick followed by = request a halant with a zero-width joiner.

This is an approved software-contract update, measured against literal, manually specified Unicode strings. The changed default and its selected tests do not establish better Nepali linguistic accuracy. The published 0.1.0-alpha.1 package is unchanged; these results describe the unpublished 0.1.0-alpha.2 source.

Input	Default / explicit full	Strict half
k / ka	क / क	क् / क
kr / kra	क्र / क्र	क्र् / क्र
kri	क्रि	क्रि
kar	कर	कर्
k/ or k followed by backtick	क्	क्
ka/ or ka followed by backtick	क्	क्
par/=yo or par followed by backtick and =yo	पर्‍यो	पर्‍यो
k^ / k~	कं / कँ	क्ं / क्ँ

The API option, explicit marks and browser typing behavior are documented in the typing reference and API guide.

The separate dated browser checks and desktop-006 machine record cover editable Roman input, explicit halves, the demo setting and installed package consumers. Their desktop integration results are separate from the core contracts below.

Recorded comparison

The machine report compares immutable baseline 987e31f7ec4e49680d7a001c078debce3d7fb326 with the new source. It pins every source file, the seed and finite mode fixtures, the tools and all historical report hashes.

Check	Prior default	New default	Explicit full	Strict half
Same revised seed contracts	130/160	160/160	160/160	—
Archived original seed contracts	142/142	128/142	—	142/142
Eighteen added seed contracts	2/18	18/18	—	—
Finite full-mode contracts	11/31	31/31	31/31	—
Finite half-mode contracts	—	—	—	31/31

On the revised seed, preferred word outputs are 101/126 → 126/126, listed candidate coverage 104/129 → 129/129, and complete text outputs 29/34 → 34/34. The before/after totals use the same revised software policy; their increase is agreement with that policy, not an accuracy gain on unchanged language labels.

Strict half also matches the archived engine's 144/144 complete raw results: 142 contract inputs and the two exploratory inputs. This comparison checks each whole word result object, including the full candidate array and ambiguous, or the whole text string. The two exploratory spelling proposals remain 0/2, retain their original labels and stay outside the contract gate.

Declared fixture revisions

The seed has 160 contracts and two exploratory cases. Fourteen prior contract rows intentionally revise their preferred output and, for words, their required singleton candidate. Each retains its ID, input, status, category, order and previous provenance, then appends a dated explanation. The other 130 of 144 prior rows remain byte-for-byte identical. Stable IDs such as half-k retain their historical names even though their current default target is full.

The revised rows are half-k, half-kr, loanword-reserved-doctor, loanword-reserved-school, loanword-check-collision, loanword-fail-collision, loanword-phail-collision, month-shift-s, month-shift-d, month-short-jan, month-short-sep, mixed-early-ordinary-text, loanword-suffix-text-boundaries and native-spelling-text-boundaries. The last three revise bare Shift T/D/S outputs inside text. The report records every old target, revised target and measured result.

Eighteen additions cover bare finals, digraphs, clusters, kar, an unsupported literal boundary, backtick/slash entry, joiner entry, nasal marks after full and explicitly half consonants, punctuation/newlines/digits and protected URL/email spans. The separate 31-case mode fixture contains literal expectations for both options without changing the general seed schema. Its targets were hand-specified from the approved typing rules before comparison; engine outputs did not generate or replace them.

All earlier dated reports and external/source-assisted references are unchanged. This record adds no human-reviewed language labels, external-corpus score, performance measurement, mobile compatibility or registry publication result. The finite compatibility checks cover their listed inputs, not every possible Roman spelling.

Reproduce

From the checkout corresponding to the report's source hashes, with Node.js 18 or later:

CONSONANT_BASELINE_DIR=$(mktemp -d)
git archive 987e31f7ec4e49680d7a001c078debce3d7fb326 \
  src package.json benchmark/cases.jsonl benchmark/run.js benchmark/reports \
  | tar -x -C "$CONSONANT_BASELINE_DIR"
node benchmark/measure-consonant-defaults.js "$CONSONANT_BASELINE_DIR" --check

npm run benchmark -- --check

Add --output <new-report.json> to save a report. An existing output file is refused. --check exits with code 1 if a revised default/full contract, finite full/half case, archived half contract or raw strict comparison fails. Invalid inputs or changed frozen fixtures/source/report identities exit with code 2. Timestamp and host details may vary; pinned fixture hashes keep denominators fixed.

The seed CLI continues to check only its default engine. The dedicated comparison creates full and half engines programmatically and runs the seed CLI against the same revised fixture to confirm its API and CLI counts agree. Replaying an earlier dated report requires its earlier implementation and fixture, rather than the current default.