Chinese support — execution backlog¶

Status: Historical work inventory (original September design; current audit linked below)
Important repository convention: bounded contributor tasks belong in campaign/tasks.json, not one GitHub issue per task. This file defines work packages and umbrella discussions; agent-ready-task-contracts.md defines how packages are split into ADR-023-compliant campaign tasks.

For the architectural dependency graph, see implementation-roadmap.md.

2026-10-29 reconciliation: The first campaign tasks are in campaign/tasks.json and some underlying QA/benchmark source has since been implemented. Inventory states may need explicit, evidence-backed reconciliation; “open” is not proof a task's code is absent, and “verified” is not proof of a real-model measurement. See as-built audit, 24 remaining task contracts, and ADR review.

1. One epic outcome¶

First-class offline Mandarin dictation¶

A user can switch an English YazSes installation to:

• Mandarin + Simplified output; or
• Mandarin + Traditional output,

use the validated product surfaces, and switch back without changing YazSes' offline-by-default architecture or regressing the default English path.

Release boundary: Mandarin only. Cantonese, arbitrary code-switching, and unvalidated surfaces remain separate.

Governing ADRs on main: v2-140 through v2-144. ADR-v2-135 now names an unrelated feature.

Done when: ADR-v2-144's evidence gate is resolved with a committed support matrix and result artifacts.

2. GitHub umbrella issues¶

These GitHub umbrella issues already exist. Use them for discussion and coordination while the governing ADRs remain Proposed:

#499 status (2026-11-06). The software parts of CHN-20..23 are on main: the English grammar sits behind the language registry and is pinned to the pre-extraction classifier by tests/fixtures/commands/en_grammar_parity.json (rule table and outputs from commit bbd285a5); the Mandarin grammar, the bounded numeral parser, Simplified and Traditional positive vectors, and a negative corpus with per-script coverage for every shipped action are tested. The user page carries the full phrase table, marked unreviewed and experimental. What remains is the native review (CMD-08), tracked in #505. Until it is recorded in zh.PHRASE_REVIEW, nothing calls Mandarin commands supported.

Do not create one GitHub issue for each 15–30 minute contributor task. ADR-023 and campaign/README.md intentionally keep those in the campaign inventory.

Current campaign mapping¶

Merged campaign PR #533 introduced the first bounded tasks from this backlog.

Open before ADR approval because they do not choose architecture:

• CHN-QA-CJK-CLEANER-001 — model-free CJK cleaner regression vectors.
• CHN-QA-HAN-SCRIPT-001 — mixed Han/ASCII/script-normalizer vectors.
• CHN-QA-MEETING-CJK-001 — minimal Meeting Mode Han-token counting reproducer.
• CHN-QA-CER-HARNESS-001 — common Mandarin CER scorer/harness (expired: the harness it targeted is no longer in this repository).
• CHN-COMPAT-X11-001, CHN-COMPAT-WAYLAND-001, CHN-COMPAT-MACOS-001, CHN-COMPAT-WINDOWS-001 — real-platform Han injection evidence.

Registered but held in verified until the common CER harness lands:

• CHN-MEASURE-WHISPER-SMALL-001
• CHN-MEASURE-WHISPER-TURBO-001
• CHN-MEASURE-WHISPER-LARGE3-001
• CHN-MEASURE-QWEN3-ASR-06B-001

Architecture-sensitive foundation and command implementation stays in the umbrella issues until ADR-v2-140..139 are accepted/superseded; do not advertise those as first-contribution tasks early.

3. Work-package register¶

These IDs are architectural work packages, not automatically campaign task IDs.

4. Milestones¶

M1 — Coherent language state¶

Packages: CHN-00, 10, 11, 12, 13, 15.

Exit criteria:

• one shared resolver;
• atomic apply;
• dry-run/no-download;
• invalid base.en + zh cannot be produced by high-level operation;
• status/doctor can explain manually invalid configs;
• English defaults unchanged.

M2 — Safe Chinese command layer¶

Packages: CHN-20, 21, 22, 23.

Exit criteria:

• English contract parity;
• safe edit/navigation Mandarin commands;
• zero false positives on curated negative fixture;
• native review of Simplified and Traditional phrases;
• terminal/open-ended commands either separately cleared or explicitly omitted.

M3 — Surface compatibility¶

Packages: CHN-30, 31F, 31M, 32A-D.

Exit criteria:

• every default-on postprocessor classified;
• file path verified;
• Meeting Mode either verified or explicitly excluded;
• each desktop platform/session has exact Han injection evidence or documented fallback.

M4 — Recognition evidence¶

Packages: CHN-40, 41, optional 42A-C.

Exit criteria:

• corpus manifest;
• scoring normalization;
• raw + requested-script CER;
• p50/p95 latency/RTF;
• RSS/load time/core-seconds;
• long-form stability;
• artifact/runtime/hardware record.

M5 — Human acceptance and support gate¶

Packages: CHN-43S, CHN-43T, CHN-50.

Exit criteria:

• separate Simplified/Traditional review;
• privacy-safe validation notes;
• risk register blockers resolved or support matrix narrowed;
• exact support wording backed by evidence.

5. Contributor lanes¶

Pure Python / no Chinese required¶

Good candidates after prerequisites merge:

• status derivation;
• profile alias tests;
• numeral parser;
• English grammar parity;
• doctor formatting/checks;
• benchmark result-schema validation;
• postprocessor code audits.

Chinese-language contributors¶

Useful without writing core Python:

• command phrase review;
• positive/negative command fixtures;
• Simplified UX terminology;
• Traditional UX terminology;
• final acceptance scenarios.

Machine translation is a drafting aid, not acceptance evidence.

Platform contributors¶

Independent tasks:

• X11;
• Wayland;
• macOS;
• Windows.

Each uses a fixed public fixture and records expected vs observed text, backend and target app. No personal clipboard/audio data.

ASR/research contributors¶

• harness;
• Whisper baseline;
• SenseVoice/Paraformer/Qwen probes;
• error analysis.

Upstream leaderboard copying is not a valid task completion.

6. Task-opening rule¶

A work package becomes an advertised campaign task only when:

1. governing ADR/API is no longer ambiguous;
2. allowed paths are exact;
3. expected time fits one sitting;
4. validation commands already exist;
5. the negative-test failure mode is named;
6. one internal implementation/review pass demonstrates the task is unambiguous;
7. risk lane is L0-L2;
8. hardware/native/model evidence tasks have cloud_agent_ready=false.

See agent-ready-task-contracts.md.

7. Common acceptance constraints¶

Every implementation task touching shared behavior inherits these:

Architecture¶

• default English model/language behavior unchanged unless a separate accepted ADR says otherwise;
• no Chinese-specific daemon/dispatch fork;
• canonical config remains source of truth;
• optional model/dependency remains lazy;
• runtime remains offline after explicit one-time artifact setup.

Tests¶

• positive case;
• negative/failure case;
• English shared-path regression where relevant;
• ordinary CI has no model/network/hardware dependency unless the task is explicitly measurement/compatibility work.

Documentation¶

• limitation stated;
• surface support matrix updated when evidence changes;
• no generic “Chinese supported” statement before ADR-v2-144.

8. Labels and metadata¶

Use repository-existing labels only. Check current labels before opening umbrellas.

Conceptual classification:

• multilingual/STT;
• commands;
• Settings/CLI;
• compatibility/platform;
• benchmark/measurement;
• localization/native review;
• help wanted / good first contribution only when ADR-023 readiness is actually satisfied.

Do not create duplicate label taxonomies just for this feature.

9. Why this structure is better for community development¶

The architectural package can stay detailed and cross-cutting, while each contributor sees only a bounded contract. That reduces:

• overlapping edits;
• giant agent-generated diffs;
• native-language judgments made by non-speakers;
• unverifiable hardware claims;
• model measurements with missing environment data;
• reviewer effort.

The roadmap is the plan. GitHub umbrella issues are the discussion layer. campaign/tasks.json is the executable contributor queue.