fix(desktop): return expected failures as codes across IPC

[orangeCatDeveloper]

Summary

Changing a task's permission mode, plan mode, orchestration or model while it is busy shows the generic "cannot change right now" toast in every locale, and an attachment that fails validation is reported by parsing an attachment_ingest:<code> token out of the error text.

Both Desktop serialization boundaries keep only Error.message. The Runtime Host already answers session.configuration.update with a structured code (session_busy, operation_conflict, …) and the Desktop client surfaces it as RuntimeHostOperationError.code — but the main-process handler rethrows, so Electron hands the renderer Error invoking remote method 'sessions:setPermissionMode': Error: Session has a pending Interaction and the presenter can only fall back.

The fix moves expected failures into return values. The five session-setting handlers answer { ok: true, session } | { ok: false, code }, narrowed to the four codes a user can act on; unexpected failures still throw. Attachment validation raises a typed AttachmentIngestBlockedError, which preload and main turn into the submitMessage / send envelope's new attachment_blocked variant. The renderer adapters unwrap a failed envelope into ExpectedOperationError(code), so the existing intent hook and sessionSettingFailureCopy keep their shape and map the code through a new updateFailures catalog. In the Runtime, the session-configuration guards throw a typed SessionConfigurationTransitionError instead of Chinese sentences.

AttachmentIngestBlockedError carries its code twice on purpose: as a field for consumers in the same process, and inside its message as the stable attachment_ingest:<code> token for anything on the far side of the IPC boundary, which drops custom Error fields. So the presenter reads the typed code where the throw is local and never parses a string.

Boundaries: the protocol is untouched, so there is no Host version skew. The plan-mode write in app-shell goes through the session-settings port rather than window.maka directly, which keeps the root-file ratchet flat.

Refs #2672

Verification

apps/desktop full dist suite        2485 pass / 0 fail
apps/desktop typecheck (4)          0 errors
renderer architecture ratchet       passes against the base, no token growth
npm run format:check                clean
protocol epoch guard                no protocol changes

Before, a session_busy answer reached the toast as the locale fallback. After:

sessionSettingFailureCopy('zh-CN', 'permission', new ExpectedOperationError('session_busy')).description
当前任务正在运行或有交互待处理，等结束后再改设置。

sessionSettingFailureCopy('en', 'permission', new ExpectedOperationError('session_busy')).description
A task is running or waiting on you. Change this setting after it settles.

Not run: Electron e2e, packaged build.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Code — traced the failure path, wrote the envelope types and the main/preload/renderer changes, the catalog copy, the tests, the rebase onto current main after #4457 merged, and this description. The decision to leave the protocol untouched and the copy wording were reviewed by hand.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No