fix(runtime): make turn acceptance retry-safe

[Neonforge98]

Problem

Exact response-loss retries were not accepted atomically across the native and external-CLI runtimes. Two native calls could both pass a read-before-upsert window, while a CLI retry could mutate model/account state or kill the live runner before discovering that the turn already existed. Project-mode submission also changed identity from the composer intent to a WorkItemRun without a durable X-to-Y receipt that callers could recover after losing the IPC response.

Solution

• Claim ordinary native and CLI turn intents with an IMMEDIATE SQLite transaction before goal, runtime, transcript, config, process, steering, or scheduler side effects.
• Return the original durable lifecycle for exact duplicates; distinguish the scheduler owner intent from a different intent that merely reuses the same client message id.
• Keep Project mode single-owned by WorkItemRun, return an explicit duplicate receipt, store originTurnIntentId, and expose the effective WorkItemRun id.
• Add agent_turn_intent_status so callers can recover status and effective identity after response loss.
• Keep legacy WorkItemRun enqueue callers and native response consumers compatible; all response fields and commands are additive.

Potential risks

• Project runs created before this change do not contain originTurnIntentId, so X-to-Y recovery applies to newly submitted turns only.
• Status lookup checks the Project run store before falling back to the ordinary session intent store; malformed duplicate Project mappings fail closed instead of guessing.
• A preparation failure closes a newly claimed intent as rejected; a deliberate retry after a returned error must use a new logical turn id.
• This PR does not migrate frontend dispatch entry points, continuation storage, cloud plane formats, or UI. Those remain separate reviewable changes.

Verification

• cargo test -p org2 exact_replay: 3 passed (initial, running, and completed retry paths)
• session-persistence concurrent claim, agent-core receipt/scheduler, and project-management idempotency/status/recovery tests: passed
• cargo check -p org2 -p agent_core -p session_persistence -p project_management --tests
• cargo clippy -p org2 -p agent_core -p session_persistence -p project_management --tests -- -D warnings
• pnpm typecheck
• targeted CliRunReceipt Vitest: 2 passed
• Pre-commit lint-staged, TypeScript, scoped clippy, and commit stats: passed

Architecture audit: no database schema migration, cloud API change, conversation-plane change, or new package-boundary dependency was introduced.

[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.