refactor(threads): strip dead v1 message writes + migrate async-research stub to v2

[tlgimenes]

Summary

Removes/migrates v1 (thread_messages) message writes now that v2 (thread_message_parts, append-only stream-of-record) is the only live write path. Two parts:

A — delete the dead v1 write path (no behavior change)

After the Phase C cutover, SqlThreadStorage.saveMessages() was reachable only via Memory.save(), which has zero callers. Removed:

• Memory.save() (api/routes/decopilot/memory.ts)
• saveMessages from the storage port + impl + org-scoped wrapper (storage/ports.ts, storage/threads.ts)
• the now-entirely-dead thread asset-hoisting proxy (object-storage/asset-hoister.ts) — its only real interception was saveMessages (kept connection/virtual-mcp hoisting)
• the dead saveMessages (upsert) test block + inert saveMessages test mocks
• stale doc-comments referencing the old saveMessagesToThread (harness packages, build-stream-request.ts, context-factory.ts)

v1 read paths (listMessages, listWithLastMessage, the v1 branches in memory.ts / list-messages.ts) are untouched — legacy v1 threads still render.

B — migrate the one live v1 writer (deep-research stub)

AsyncResearchJobs.markPolling() wrote a stub assistant message to thread_messages unconditionally so a browser refresh mid-poll can reconstruct the in-flight tool call. But v2 threads read thread_message_parts, so on v2 the stub went to a table nobody reads (refresh-recovery silently broken: tool-input-available parts are non-final → never persisted by the projector, and v2 loadWindow only surfaces messages with a finish anchor).

Now version-gated:

• v2 thread → seed a tool_call part + finish anchor into thread_message_parts (run_id == threadId, deterministic ids, idempotent) so loadWindow surfaces it
• legacy v1 thread → unchanged thread_messages stub
• stub cleanup (deleteStubMessage on terminal transitions + sweepAbandoned) now clears both tables (best-effort)

Testing

• bun run check — passes across all workspaces
• bun run lint — 0 errors (3 pre-existing warnings in untouched files)
• New storage/async-research-jobs.integration.test.ts (real PG): v2 markPolling seeds a folded, loadable assistant message with the tool-web_search input-available part; markCompleted removes it; v1 path still writes thread_messages.
• Affected storage tests: 18 pass, 0 fail.

Follow-ups (not in this PR)

• listWithLastMessage ("Suggested actions") reads thread_messages only — has no v2 equivalent (latent v2 gap).
• Eventual v1 read removal would need a thread_messages → thread_message_parts backfill.

🤖 Generated with Claude Code

---

Summary by cubic

Removes dead v1 message writes and migrates the async-research polling stub to v2 thread_message_parts, fixing refresh recovery on v2 threads and simplifying storage.

• Refactors

  • Removed Memory.save() and ThreadStoragePort.saveMessages (and all impls/wrappers).
  • Dropped the thread message asset-hoisting proxy; kept hoisting for connections and virtualMcps only.
  • Deleted obsolete tests and comments referencing saveMessages/inline persistence.
  • v1 read paths remain (legacy threads still render).

• Bug Fixes

  • AsyncResearchJobs.markPolling() now seeds a v2 stub into thread_message_parts with a tool_call part and finish anchor; v1 threads unchanged.
  • Terminal cleanup and abandoned sweep delete stubs from both thread_messages and thread_message_parts.
  • Added PG integration tests covering v2 seed/cleanup and ensuring no v1 writes on v2 threads.

^{Written for commit 9f6dd95. Summary will update on new commits.}