refactor(conversation): centralize execution registry

[Neonforge98]

Problem

Shared-conversation one-shot runners persist their hidden-session registry directly inside conversationTurnRunner as one global localStorage blob keyed with colon-concatenated organization and session ids. That couples execution to sidebar policy, permits key collisions, can lose an unrelated conversation update from another desktop window, and does not isolate records across cloud accounts.

Solution

• Add one canonical local conversation execution store with a separate entry per (executor scope, root session) tuple.
• Scope new records by signed-in cloud identity plus organization, while preserving the root session as the second identity dimension.
• Move runner hiding and lifecycle registration behind a small conversationRunnerSessions facade.
• Keep reading the shipped legacy one-shot registry so upgrades never reveal plumbing sessions in My Sessions.
• Record proven terminal runners for later safe cleanup without adding continuation behavior in this PR.
• Leave Team Chat, audience routing, cloud schemas, Work Item dispatch, and the current one-shot execution semantics unchanged.

Potential risks

• This PR is intentionally stacked on refactor(runtime): route turn producers through dispatcher #942 and should be reviewed or merged after fix(runtime): make turn acceptance retry-safe #940, feat(runtime): add canonical turn dispatch boundary #941, and refactor(runtime): route turn producers through dispatcher #942.
• A failed localStorage write remains best-effort, matching current behavior; the visible failure mode is only a local plumbing row appearing in My Sessions.
• Legacy records are retained and read rather than destructively rewritten because the old key lacks account identity.

Validation

• pnpm typecheck
• ESLint and Prettier on all 6 changed files
• 3 focused Vitest files, 29 tests passed
• Conversation execution-store suite: 7 tests covering account isolation, tuple collisions, legacy visibility, malformed data, nested window writes, lifecycle, and cleanup
• Madge scan from all changed entry points: 1220 files, no circular dependency
• Repository commit hooks passed

[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Superseded by the final two-PR design: #844 removes the Work Item-specific remote continuation bridge, and stacked #939 keeps the reusable audience routing plus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.