Mixed-text benchmark record — 2026-09-27

This record measures whether the default whole-text converter follows the selected mixed-text preservation contract. The API guide defines the current address policy, punctuation boundaries, and the engine opt-out. This dated record covers the first address-preservation policy; the early-address follow-up records later cue preservation separately and does not replace these historical outputs. The machine report records exact inputs and outputs, per-set counts, engine and fixture identities, changed case IDs, commands, and limitations.

These are project-authored software behavior contracts. No independently reviewed linguistic labels were added, and the result is not a Nepali accuracy, timing, or browser-compatibility score.

Recorded results

Recorded on 2026-09-27 using Node.js v22.22.3 on Darwin arm64. The baseline is commit 4e6cc4eed855f0d97bff366c619e5f4eff50e2a8, extracted with git archive. The after run used a modified working tree based on that commit; its core files are pinned by SHA-256 in the report rather than assigned an uncreated commit identity. That exact after engine was subsequently committed as c0679b5e41a56eb5bab762a422243a400adefe1f, which the reproduction commands below archive along with its original fixture and report.

Check	Before	After	Interpretation
Original contracts	109/109	109/109	Previous word and text contracts preserved
Same expanded contracts	110/119	119/119	Both engines use the same frozen fixture
Expanded word top output	102/102	102/102	Word behavior unchanged
Expanded text exact output	8/17	17/17	Selected complete strings, including addresses and punctuation
Added mixed-text contracts	1/10	10/10	Nine preservation targets change; the ordinary-mark control already passed
Exploratory proposals	0/2	0/2	Unreviewed targets; excluded from the contract gate

The expanded fixture has 121 rows: 119 contracts and two exploratory cases. All 111 previous rows and IDs remain byte-for-byte unchanged. The added cases are:

Case ID	Behavior covered
mixed-url-case-port	Original HTTP(S) host/path case and numeric port
mixed-email-tag	Email case, dotted local part, plus tag, and danda outside the address
mixed-www-path	A www. address and path beside a Nepali word
mixed-bare-domain-loanword	A bare subdomain beside automatically converted loanwords
mixed-query-explicit-marks	^, ~, /=, and `
mixed-path-wrappers	Enclosing punctuation, balanced path parentheses, and an email
mixed-unicode-boundary	ASCII addresses beside existing Devanagari and emoji
mixed-dotted-roman-shape	Documented heuristic: domain-shaped pani.paani stays literal
mixed-ordinary-marks-and-decimal	Existing decimal, explicit marks, and conversion of unprotected identifiers
mixed-month-domain	Month aliases beside a domain and mailto: address

For example, camera camera.co.np phone changes from क्यामेरा क्यामेरा.चो.न्प् फोन to क्यामेरा camera.co.np फोन. The unchanged control 3.14 par/=yo ka^ kaa~ camera_file camera123| still gives 3.14 पर्‍यो कं काँ क्यामेरा_फाइल क्यामेरा123।: code-like identifiers do not acquire general literal preservation.

Engine and fixture identity

The report pins src/index.js, src/index.d.ts, src/text-policy.js, src/lexicon.js, and src/phonetic.js. The baseline has no text-policy module; its digest is recorded as null. Lexicon and phonetic file digests are identical before and after. The expanded fixture SHA-256 is 3e2674bc1bc9c0acd58fcc159eecd6f4666b94897761d33512c76f2fd8addb16; the original fixture digest is 2ed3e1aba19bfaaf2c48fa8b07f9dac9de3639ced2c6048648c206d486aa99ac.

The benchmark runner digest is also recorded. A measurement-script digest identifies the script used for the recorded comparison, but it is not a shipped runner or a requirement to reconstruct its temporary paths. The portable commands below run the repository runner with frozen fixtures and reject different core, runner, or fixture identities.

Reproduce the recorded comparison

Use Node.js 18 or later and a repository checkout whose Git history contains both immutable commits below. Run from the repository root. Both engines, the frozen 119-contract fixture, and the recorded report are extracted from those commits, so subsequent working-tree changes do not change this dated comparison. Hash checks reject different archived content.

MIXED_BASELINE_DIR=$(mktemp -d)
MIXED_AFTER_DIR=$(mktemp -d)
MIXED_FIXTURE_FILE=$(mktemp)
git archive 4e6cc4eed855f0d97bff366c619e5f4eff50e2a8 | tar -x -C "$MIXED_BASELINE_DIR"
git archive c0679b5e41a56eb5bab762a422243a400adefe1f | tar -x -C "$MIXED_AFTER_DIR"
cp "$MIXED_AFTER_DIR/benchmark/cases.jsonl" "$MIXED_FIXTURE_FILE"
node --input-type=module - "$MIXED_BASELINE_DIR" "$MIXED_AFTER_DIR" "$MIXED_FIXTURE_FILE" <<'JS'
import { readFile } from 'node:fs/promises';
import { createHash } from 'node:crypto';
const report = JSON.parse(await readFile(process.argv[3] + '/benchmark/reports/mixed-text-001.json', 'utf8'));
const hash = async (path) => createHash('sha256').update(await readFile(path)).digest('hex');
for (const [root, identity] of [[process.argv[2], report.baseline], [process.argv[3], report.after]]) {
  for (const [file, expected] of Object.entries(identity.coreFileSha256)) {
    if (expected === null) {
      try {
        await readFile(root + '/' + file);
        throw new Error('Unexpected baseline file: ' + file);
      } catch (error) {
        if (error.code !== 'ENOENT') throw error;
      }
    } else if (await hash(root + '/' + file) !== expected) {
      throw new Error('Core differs: ' + file);
    }
  }
  if (await hash(root + '/benchmark/run.js') !== report.benchmarkToolSha256) {
    throw new Error('Benchmark runner differs');
  }
}
if (await hash(process.argv[4]) !== report.fixture.sha256) {
  throw new Error('Recorded mixed-text fixture differs');
}
if (await hash(process.argv[2] + '/benchmark/cases.jsonl') !== report.fixture.priorFixtureSha256) {
  throw new Error('Original fixture differs');
}
JS

Run these benchmark commands separately if the shell stops on a nonzero exit status:

node "$MIXED_BASELINE_DIR/benchmark/run.js" --fixtures "$MIXED_FIXTURE_FILE" --check
node "$MIXED_AFTER_DIR/benchmark/run.js" --fixtures "$MIXED_FIXTURE_FILE" --check

The baseline intentionally exits with code 1 for 110/119 contracts; the recorded after engine exits with code 0 for 119/119. Check the original denominator separately:

node "$MIXED_BASELINE_DIR/benchmark/run.js" --fixtures "$MIXED_BASELINE_DIR/benchmark/cases.jsonl" --check
node "$MIXED_AFTER_DIR/benchmark/run.js" --fixtures "$MIXED_BASELINE_DIR/benchmark/cases.jsonl" --check

Both commands pass 109/109 original contracts. Exploratory rows are printed but do not fail the gate. For the latest working tree rather than this dated comparison, run npm run benchmark -- --check; use npm test for the separate functional suite.

Scope limits

The policy infers recognizable ASCII host and email shapes without DNS, public-suffix validation, full URL/email validation, or network access. Domain-shaped Roman words can therefore be preserved. Unrecognized code, filenames, arbitrary English, and Unicode hostnames/mailboxes have no general literal-text guarantee. See the typing reference for explicit English controls.

This report measures default convertText only. Opt-out behavior, custom-entry precedence, the unchanged convertWord API, and browser live-editing paths are separate functional-test concerns. No new external linguistic comparison, independently reviewed corpus labels, mobile/browser support claim, assistive-technology result, or performance result is established. The earlier month, loanword, and ra-ya results retain their original dated counts and evidence.

Early-address follow-up — 2026-09-27

The early-address machine report records a separate correction prompted by incomplete addresses appearing as Nepali until enough of the host had been typed. The shared scanner now preserves unfinished addresses at http:, https:, www., an ordinary ASCII local part followed by @, or the first letter after a domain dot. Plain camera, sentence-final camera., decimals, and explicit Nepali marks keep their previous behavior. It does not predict an address before a cue appears or validate that the address is real.

The baseline is the immutable first mixed-text implementation, c0679b5e41a56eb5bab762a422243a400adefe1f. The corrected after engine is identified by core-file hashes in the report. Both runs use the same expanded fixture with 127 rows: 125 contracts and two excluded exploratory proposals. All 121 prior rows remain byte-for-byte unchanged.

Check	Before	After
Prior mixed-text contracts	119/119	119/119
Same expanded contracts	120/125	125/125
Word top output	102/102	102/102
Text exact output	18/23	23/23
Six early-address additions	1/6	6/6
Exploratory proposals, excluded from gate	0/2	0/2

The six appended IDs are mixed-early-http-scheme, mixed-early-http-userinfo, mixed-early-www, mixed-early-email, mixed-early-domain, and mixed-early-ordinary-text. Five selected strings intentionally change output; the ordinary-text control already passed. Separate functional tests cover intermediate keystrokes, cue deletion, caret movement, undo/redo, native input, paste, and composition. This default convertText comparison adds no independent linguistic labels, external word evaluation, performance measurement, or browser/device certification.

Reproduce the early-address comparison

Use Node.js 18 or later and archive both immutable engines below. Commit 995909554efa84691f5a481a0d27def58e308d2f contains the corrected after engine, frozen fixture, and report matching this dated record. Hash checks stop the comparison if the selected sources differ. The earlier 119-contract comparison remains reproducible from its own immutable snapshots above.

EARLY_BASELINE_DIR=$(mktemp -d)
EARLY_AFTER_DIR=$(mktemp -d)
EARLY_FIXTURE_FILE=$(mktemp)
git archive c0679b5e41a56eb5bab762a422243a400adefe1f | tar -x -C "$EARLY_BASELINE_DIR"
git archive 995909554efa84691f5a481a0d27def58e308d2f | tar -x -C "$EARLY_AFTER_DIR"
cp "$EARLY_AFTER_DIR/benchmark/cases.jsonl" "$EARLY_FIXTURE_FILE"
node --input-type=module - "$EARLY_BASELINE_DIR" "$EARLY_AFTER_DIR" "$EARLY_FIXTURE_FILE" <<'JS'
import { readFile } from 'node:fs/promises';
import { createHash } from 'node:crypto';
const report = JSON.parse(await readFile(process.argv[3] + '/benchmark/reports/early-address-001.json', 'utf8'));
const hash = async (path) => createHash('sha256').update(await readFile(path)).digest('hex');
for (const [root, identity] of [[process.argv[2], report.baseline], [process.argv[3], report.after]]) {
  for (const [file, expected] of Object.entries(identity.coreFileSha256)) {
    if (await hash(root + '/' + file) !== expected) throw new Error('Core differs: ' + file);
  }
  if (await hash(root + '/benchmark/run.js') !== report.benchmarkToolSha256) {
    throw new Error('Benchmark runner differs');
  }
}
if (await hash(process.argv[4]) !== report.fixture.sha256) throw new Error('Expanded fixture differs');
if (await hash(process.argv[2] + '/benchmark/cases.jsonl') !== report.fixture.priorFixtureSha256) {
  throw new Error('Prior mixed-text fixture differs');
}
JS

Run the commands separately if the shell stops on a nonzero exit status:

node "$EARLY_BASELINE_DIR/benchmark/run.js" --fixtures "$EARLY_FIXTURE_FILE" --check
node "$EARLY_AFTER_DIR/benchmark/run.js" --fixtures "$EARLY_FIXTURE_FILE" --check

The baseline intentionally exits 1 for 120/125 contracts; the matching after engine exits 0 for 125/125. Check the unchanged prior denominator with:

node "$EARLY_BASELINE_DIR/benchmark/run.js" --fixtures "$EARLY_BASELINE_DIR/benchmark/cases.jsonl" --check
node "$EARLY_AFTER_DIR/benchmark/run.js" --fixtures "$EARLY_BASELINE_DIR/benchmark/cases.jsonl" --check

Both pass 119/119 prior contracts. These two dated reports retain distinct fixture and engine identities; use npm run benchmark -- --check for the latest checkout rather than a historical result.