refactor(i18n): track cross-surface i18n and l10n governance

[limityan]

Background

BitFun aligned i18n and l10n across Web UI, mobile-web, installer, backend, and relay static pages around one shared language contract. Each product surface still owns and loads only the resources it needs.

Completion Status

This refactor stream is complete after the active closure PR is merged. Future i18n drift should be handled as ordinary maintenance using the documented architecture, audit reports, and lightweight verification rules.

Final Architecture Snapshot

• Canonical locale identity, aliases, fallback chains, surface defaults, and generated contracts are centralized.
• Product surfaces keep independent resource ownership and do not import another surface's full locale catalog.
• Shared terms are reserved for stable cross-surface concepts such as product names, feature names, tool names, modes, connection methods, and status labels.
• Context-specific page, workflow, toast, validation, and feature copy stays in the nearest owner namespace.
• Static keys, dynamic keys, placeholder parity, literal fallbacks, direct locale formatting, shared-term duplicates, same-text locale inventory, and actionable l10n signals are audited as separate categories.
• Local prechecks stay lightweight; broader platform/build coverage remains protected by CI.

Final Progress Snapshot

• Unified locale contract and generated surface contracts are in place across frontend/backend surfaces.
• Web UI keeps bootstrap namespaces separate from lazy locale namespaces.
• mobile-web, installer, backend, and relay static pages have their own resource paths and generated contract entry points.
• Shared terms are available across surfaces and are used for stable mode, agent, tool, connection-method, product, feature, and Deep Review labels without forcing cross-surface bundle loading.
• Dynamic-key candidates are explicitly tracked with source-proof metadata for known dynamic access paths.
• Static Web UI literal fallback debt and direct locale-formatting candidates are zero.
• Broad zh-CN/zh-TW same-text copy is visible as non-blocking review inventory; actionable l10n signal candidates stay separate at zero.
• Durable i18n docs cover resource ownership, lookup rules, audit semantics, and lightweight verification.

Final Key Metrics

Baseline is the earliest tracked governance snapshot for each comparable metric in this refactor stream.

Metric	Baseline	Final	Change / meaning
Literal fallback candidates	457	0	-457
Direct locale-format candidates	32	0	-32
Shared-term duplicate candidates	214	185	-29; remaining entries are governed no-growth review data
Dynamic key candidates	194	193	-1; all current dynamic candidates are source-proven
Dynamic candidates with source proof	0	193	+193
Actionable same-text l10n signal candidates	1	0	-1 actual signal fixed
Same-text locale review inventory	not reported	790	Non-blocking copy-review inventory; not a pass/fail metric
Confirmed unused keys	0	0	0

Post-Close Guardrails

• Keep source-proof metadata current for dynamic key access.
• Migrate shared-term duplicates only when the local value is the same stable concept; keep context-sensitive copy local.
• Treat same-text locale inventory as copy-review input, not deletion evidence or a pass/fail quality metric.
• Keep static defaultValue fallback and direct locale-format candidates at zero by repairing locale resources or using surface i18n helpers.
• Adding a future locale or product surface must update generated contracts, resource boundaries, fallback rules, and audit coverage together.

Done Criteria

• Adding a locale or product surface has clear contract, resource ownership, fallback, generation, and audit entry points.
• Product surfaces do not load resource catalogs they do not own.
• Key deletion can rule out static and dynamic references with auditable evidence.
• Stable concept labels have an authoritative source, repeated values have a reviewable report, and l10n signal issues are tracked separately from structural i18n cleanup.
• Durable docs are enough for future contributors to add, read, migrate, and audit i18n resources consistently.

[limityan]

Update from #972 (refactor(i18n): tighten shared terms and dynamic key proofs).

Key Comparison

Metric	Before	After	Delta
Shared-term duplicate candidates	214	211	-3
Installer shared-term duplicates	3	0	-3
product.name duplicate candidates	6	3	-3
Dynamic key candidates	194	194	0
Dynamic candidates with source references	0	194	+194

Notes:

• Installer titlebar defaults now read shared.product.name; copied titlebar.default keys were removed.
• Dynamic-key contracts now require audited sourceReferences, so future cleanup can distinguish documented dynamic access from removable keys.

[limityan]

Follow-up snapshot for PR #977.

This PR continues the i18n/l10n governance track without closing this issue. It adds audit visibility for literal fallback and direct locale-formatting debt, lowers several existing debt baselines, keeps remaining work visible for later cleanup, and adds a review follow-up so direct Intl.* calls without new cannot bypass the locale-format audit.

Metric	Before	After #977	Delta
Confirmed unused keys	0	0	0
Dynamic key candidates	194	193	-1
Shared-term duplicate candidates	206	201	-5
L10n quality candidates	1093	1087	-6
Literal defaultValue fallbacks	457	396	-61
Direct locale-format calls	32	24	-8

Review/CI follow-up:

• Fixed Web UI test mocks that still depended on literal fallback behavior after static fallback strings were removed.
• Hardened the locale-format audit to catch both new Intl.* and bare Intl.* calls, plus existing toLocale* calls.
• GitHub Actions CI passed on a3c966bf: Frontend Build and all Rust Build Check jobs for ubuntu-latest, macos-15, and windows-latest.
• Local validation includes i18n contract/audit/report, Web UI full tests, Web/Mobile type checks, Web/Mobile builds, repo hygiene, GitHub config, and diff hygiene.

Remaining work is still tracked here rather than tied to PR closure: continue reducing Web UI literal fallbacks, review high-confidence zh-TW localization candidates, migrate only semantically stable shared concept labels, and move the remaining direct locale-formatting call sites behind surface i18n helpers.

[limityan]

Update: PR #990 continues the i18n/l10n cleanup stream and combines the planned PR A + PR B scope into one reviewable PR.

Functional snapshot:

• Web UI flow-chat/Deep Review literal fallback debt was reduced where keys already exist in locale resources.
• Web UI and Mobile Web date/number formatting now routes through owned i18n formatting helpers at product call sites.
• zh-TW high-confidence terminology was improved across Web UI, Mobile Web, installer, and core resources while keeping locale files readable UTF-8.
• Governance reporting now keeps zero-count surfaces visible, which makes future surface regressions easier to detect.
• Remaining direct locale-format debt is isolated to two core miniapp assets and stays tracked by the baseline.

Metric	Before	After	Change
Literal fallback candidates	396	149	-247
Direct locale-format candidates	24	2	-22
l10n quality candidates	1087	768	-319
Shared-term duplicate candidates	201	201	0
Mobile Web l10n candidates	19	0	-19

Verification for PR #990 includes i18n audit/contract, Web UI type-check/lint/tests, Mobile Web type-check, and repo hygiene. Heavy installer/desktop builds were left to CI for this i18n-focused slice.

[limityan]

P0 follow-up: #995 restores the latest i18n governance snapshot after the zh-TW write-tool-mode label regression.

Metric	Before	After	Change
l10n quality candidates	769	768	-1
Web UI l10n candidates	726	725	-1
settings/agentic-tools l10n candidates	6	5	-1

This keeps the top-level issue snapshot unchanged at the intended 768 l10n-quality-candidate state.