feat(conversations): select local execution runtime

[Neonforge98]

Problem

A member starting a local continuation can currently use only the native Agent runtime, even when a managed External CLI is installed. Durable Work Item redelivery also carries a prepared runner id but the frontend does not fence recovery to that exact local Session.

Solution

Separate assigned Agent identity from the local execution runtime in the shared setup dialog. Conversation runners may choose Native Agent or an installed managed External CLI that supports GUI/headless execution; pure TUI clients are excluded. Both choices still use the same SessionService create/send/wait/push path, persistent continuation record, cursor, overlay, and cleanup logic. Persist the CLI platform with the continuation and require a redelivered prepared run to recover the exact same runner and compatible Agent fingerprint. Work Item assigned Agents are locked in the dialog. Team Chat remains ordinary human chat.

Potential risks

External CLI execution depends on the existing managed GUI transport and local CLI authentication; a removed CLI causes remembered setup to be cleared and re-prompted through the existing fallback. If the exact prepared local Session or its continuation record is missing, the durable run stays retryable instead of silently switching runner identity. Replay-copy forks remain native-only and reject a programmatic CLI selection.

Validation

• 67 focused Vitest tests passed
• pnpm exec tsc --noEmit --pretty false
• ESLint passed on touched TypeScript files
• collaboration.session i18n parity passed for all 13 locales
• Normal pre-commit hooks passed

Stacked on #947.

[Neonforge98]

Superseded by the final two-PR design: #844 removes the Work Item-specific remote continuation bridge, and stacked #939 keeps the reusable audience routing plus sender-local provider-native continuation. This dispatcher/store/handoff stack is intentionally retired.