#663 Implement ephemeral data as in-memory repository

Problem

Swamp defines an "ephemeral" lifetime for data artifacts — data that should only exist for the duration of a method invocation or workflow execution. Today, "ephemeral" is accepted by the type system and schema but not implemented. The lifecycle service logs a warning and returns null, and ephemeral data gets written to disk and never cleaned up — behaving identically to "infinite".

Proposed Solution

Implement ephemeral data as a global in-memory store that:

1. Never writes to disk — content and metadata live only in memory
2. Is indexed for search — uses an in-memory SQLite catalog (:memory:) so it behaves identically to persistent data from a query perspective (data.latest(), data.query(), data.search() all work transparently)
3. Is scoped to execution lifetime — when the workflow run or method run that created it completes, the in-memory store is dropped
4. Is available across remote execution — workers dispatched by the orchestrator read/write ephemeral data through the same data plane interfaces
5. Requires no new CEL functions — existing data.latest() transparently resolves from the in-memory store; the engine determines whether data is in-memory or on disk

Why

Ephemeral data is the foundation for dynamic data lookups — intermediate results, computed context, and runtime-resolved references that models and workflows produce and consume within a single execution. Without this, there is no way to pass transient computed data between workflow steps without polluting the persistent datastore.

Scope of Changes

The implementation involves these new components:

• InMemoryUnifiedDataRepository (src/infrastructure/persistence/in_memory_data_repository.ts) — full implementation of the UnifiedDataRepository interface using Map-based storage, with a :memory: CatalogStore for search indexing
• CompositeUnifiedDataRepository (src/domain/data/composite_data_repository.ts) — routes writes by lifetime === "ephemeral" to the in-memory repo, reads check in-memory first then fall back to the filesystem repo
• CompositeDataQueryService (src/domain/data/composite_data_query_service.ts) — merges query results from both persistent and ephemeral catalogs so CEL queries work transparently

And modifications to existing wiring:

• DefaultStepExecutor.buildDeps() — accept and wire ephemeral repo
• WorkflowExecutionService.run() — create ephemeral store at workflow start, dispose in finally blocks
• modelMethodRun() — same pattern for standalone method runs
• executeWorkflowWithLocks() — serve-side creation/disposal
• DataLifecycleService — remove the "not implemented" warning for ephemeral

A detailed design document is at design/ephemeral-data.md with exact file paths, line numbers, interface contracts, and a phased implementation plan.