[Astro-Han]

Summary

Part of #860. Adds the Xiaomi MiMo Token Plan subscription as a three-region cohort under three independent persisted provider ids:

• xiaomi-token-plan-cn — Xiaomi Token Plan (China), https://token-plan-cn.xiaomimimo.com/v1, category domestic
• xiaomi-token-plan-sgp — Xiaomi Token Plan (Singapore), https://token-plan-sgp.xiaomimimo.com/v1, category overseas
• xiaomi-token-plan-ams — Xiaomi Token Plan (Europe), https://token-plan-ams.xiaomimimo.com/v1, category overseas

Each is a registry-driven plans-group provider: protocol: 'openai', runtimeAdapter: { kind: 'openai-compatible', name: 'provider' }, authKind: 'api_key', catalogBadge: 'Token', modelDiscovery: { kind: 'fallback' } (the Token Plan /v1 endpoint publishes no /models list). Brand marks reuse the already-vendored xiaomimimo.svg (no new SVG, no THIRD_PARTY_LICENSES change).

fallbackModels = ['mimo-v2.5-pro', 'mimo-v2.5'] for all three regions. This is a hard-coded, guard-asserted allowlist. models.dev's snapshot for these ids still carries the deprecated mimo-v2-pro (2026-06-30) and speech-only mimo-v2-tts / mimo-v2.5-tts* models; the guard asserts functionCalling on the two documented MiMo chat models and the allowlist deliberately excludes every deprecated and TTS/ASR model.

License policy (interactive-only): Token Plan is officially coding-only and prohibits API automation. Following the tencent-token-plan precedent (#914), no packages/headless/src/provider-env.ts entry is added; headless credential loading and Harbor cell resolution both resolve apiKey === '' even when an env key and a credentials file are present, asserted by new boundary tests.

Why one PR for three regions

The three regions are identical along every dimension except hostname — same protocol, same /v1 OpenAI-compatible path, same tp- key namespace, same model set, same coding-only license, same fallback/discovery behavior. Splitting would produce three near-duplicate reviews with no independent build/test/review boundary, so they ship as one cohort.

Deviations from the playbook

• Added apps/desktop/src/renderer/settings/provider-display-copy.ts entries for the three ids. This file is a required exhaustive Record<ProviderType, ...>; the compiler flagged it once the registry keys were added. The playbook's file inventory did not list it, but it is mandatory wiring, not gold-plating.
• The generated packages/core/src/model-metadata.generated.ts was regenerated via npm run sync:model-metadata, then reduced to only the three new provider blocks (plus the three type-union updates). The live models.dev fetch also carried unrelated drift for other providers (ollama-cloud, openrouter, etc.); that drift was excluded to keep this PR scoped and avoid conflicts with the parallel Alibaba slices.
• No packages/runtime/src/model-factory.ts buildProviderOptions case was added. The existing xiaomi direct provider uses the same MiMo models with no reasoning-effort wiring and passes the conformance matrix; the runtime gap report is clean for all three new ids, so no case or override is warranted.

Verification

Run from the worktree; all deterministic, on the inert defaults (no live provider calls).

• npm run typecheck — pass (all workspaces)
• npm run test:scripts — 109 pass / 0 fail (includes the new Xiaomi Token Plan sync test)
• npm --workspace @maka/core test — 1043 pass / 0 fail
• npm --workspace @maka/runtime test — 1974 pass / 0 fail / 7 skipped (conformance matrix gap report clean for the three new ids)
• npm --workspace @maka/storage test — 284 pass / 0 fail / 1 skipped
• npm --workspace @maka/headless test — 1010 pass / 0 fail / 1 skipped
• npm --workspace maka-agent test — 409 pass / 0 fail
• npm --workspace @maka/desktop test — 2556 pass / 0 fail
• npm run check:stale — dist is fresh
• git diff --check — clean

No sanitized live smoke was run: maintainers have no Token Plan subscription, so all coverage is from deterministic tests. Unconfirmed items from the facts research, to be corrected on user feedback:

• Auth header: official examples use an api-key header while models.dev marks the ids openai-compatible (implying Bearer). This PR sends Bearer to match the existing xiaomi direct provider. Not empirically verified.
• GET /models: the Token Plan /v1 endpoint's discovery behavior is undocumented and unverified; modeled as fallback (no /models request), matching the coding-only interactive contract.