[Feature] mcode provider add should support context and output limits for custom/local providers #161

Problem

I configure a local llama.cpp server as a custom provider with:

mcode provider add --name my-provider --base-url https://example.com/v1 \
  --api-format openai-completions --model my-model \
  --api-key-env MCODE_PROVIDER_API_KEY --use

There is no CLI way to set the model's context window (ctx) or output limit at creation time.

This matters because model discovery (ModelDiscoveryClient in packages/local-runtime-v2/src/service/model-system/connectivity/discover-models.ts) only discovers modelId/displayName from GET /v1/models or /models — it does not probe num_ctx / n_ctx / context_length. At resolve time planCustomProviderResolution() in packages/local-runtime-v2/src/service/model-system/resolution/model-resolver-byok.ts falls back to BYOK_FALLBACK_MODEL_LIMITS (200_000 ctx / 16_384 output) when modelConfig.limit is unset. So a local model started with e.g. --ctx-size 32768 is silently resolved with the wrong window unless the user edits config.yaml manually.

Current chain

• packages/tui/src/cli/program.ts (provider add): only --name, --base-url, --api-format, --model <id> (repeatable string[]), --api-key-env, --use.
• packages/tui/src/cli/provider-command.ts (McodeProviderCliRequest): models: readonly string[], mapped as models.map((modelId) => ({ modelId })).
• Downstream already supports limits: McodeProviderModelInput.limit?: { context?: number; output?: number } in packages/tui/src/provider/contract.ts, consumed by modelsFromInputs() / normalizeModelLimit() in packages/local-runtime-v2/src/service/model-system/management/service-input.ts and customProviderLimits() in model-resolver-byok.ts.

So this is a CLI-surface gap only.

Proposal

Extend mcode provider add with e.g.:

mcode provider add --name my-provider --base-url http://localhost:8080/v1 \
  --api-format openai-completions --model my-model \
  --api-key-env MCODE_PROVIDER_API_KEY \
  --context-limit 32768 --output-limit 4096 --use

• Validate as positive safe integers, consistent with existing normalizeModelLimit() / updateLocalModelSelection() validation.
• Document semantics for repeatable --model: simplest is "same limits apply to all listed models".
• Consider whether provider update / existing updateUserModelParameters() path also needs CLI exposure, or if add-time support is sufficient for v1.
• Update README.md + README_ZH.md + docs/examples.md (English primary per CONTRIBUTING.md).

Acceptance

• mcode provider add ... --context-limit N --output-limit M persists custom_provider.<key>.models.<id>.limit = { context: N, output: M } in <dataDir>/config.yaml.
• mcode provider list --json / provider view shows the configured limits.
• Invalid (zero/negative/non-integer) values are rejected with a clear error.
• Existing behavior without the new flags is unchanged.