feat(conversations): unify native session continuation

[Neonforge98]

Problem

Imported, Team, Codex, and Claude Code sessions could be displayed, but continuation was not one generic provider-native conversation flow. Runtime changes could lose native role and tool structure, interrupted partial work, queue semantics, sender identity, or native-app discoverability. Team Chat and failed-message UX also followed separate paths.

Solution

Add one provider-neutral canonical conversation event plane and materialize that history into each target provider native role and tool transcript. Same-runtime continuation reuses the provider native conversation; a runtime switch creates or updates the target provider conversation under the same ORG2 canonical root. User, assistant, tool-call, tool-result, compact, interrupted-partial, Team Chat, and audience events remain structured and are never flattened into one transcript prompt.

The composer reuses the existing durable message queue and the Agent or Team Chat control from PR #1127. turnIntentId is the end-to-end identity for optimistic, queued, accepted, failed, retried, and provider-persisted rows. Pending and failed messages remain visible in the transcript. Existing chat viewport ownership from PR #843 remains the only follow and scroll implementation.

Provider adapters publish Codex catalog entries and Claude transcript plus metadata atomically. Context exhaustion is handled only after the provider reports it, by rebuilding a valid provider-native message list and retrying once. Work Items may trigger this capability later, but do not own a separate continuation implementation.

This adds no cloud host, sender-device dispatcher, checkpoint transport, E2EE transcript subsystem, second queue, or second scroll controller.

Potential risks

Native provider formats and desktop-app caches can change upstream. The adapters validate fidelity, fail closed for unsupported content, isolate provider-specific catalog logic, and publish transactionally. Runtime switching also depends on strict ordering and idempotency around queued or interrupted turns.

Live Claude verification is currently blocked on this machine because the persisted Anthropic-1 OAuth refresh token is invalid. The independent cloud dual-instance spec is also skipped without E2E_CLOUD service credentials. Those paths are not claimed as passed.

Verification

• Stacked on PR feat(pm): expand MCP, routines, work items, inbox, and skills #844 at 1eaec91; both stack bases resolve directly to develop at 68e7aff.
• Complete Vitest suite passed: 1,358 files and 10,747 tests.
• pnpm typecheck, pnpm typecheck:fast, and pnpm lint:fast passed.
• Targeted queue and canonical transcript regression suite passed: 38 tests.
• Rust native materializer passed 30 tests; Codex native catalog passed 5 tests. Commit hooks also passed Cargo clippy.
• Real Codex App catalog lookup returned the migrated UUID, title, and cwd. The App was navigated to UUID 01a05745-9ca7-7702-8f33-9f9707d24499 and rendered four native turns, including repository exploration and the Codex to Claude to Codex return increment.
• Claude Desktop visibly lists the migrated sessions, but click-open automation is blocked by the local Computer Use window error and the live Claude CLI cannot authenticate with the expired OAuth session.
• Queue audit found no enqueueCountAtom, queueFlushRequestAtom, fixed visibility delay, text-based synthetic suppression, second queue, or second follow and scroll implementation.
• The cloud dual-instance E2E launches successfully but skips without E2E_CLOUD credentials; this remains a required live acceptance check.