A stream end always records the run's status (CP-87)

This release makes the end status of every agent stream explicit. Before it, StreamManager.endStream's terminal argument was optional, and every reader that found no recorded status guessed completed. A run that was interrupted (or, on DBOS, a cancelled workflow's failed run) could therefore reach a cross-service parent as end{completed} with null output, and the parent folded it as a successful child.

It is a minor release (a 0.x package: minor marks an observable behaviour change) of @helix-agents/core, @helix-agents/store-memory, @helix-agents/store-redis, @helix-agents/store-cloudflare, @helix-agents/runtime-js, @helix-agents/runtime-cloudflare, @helix-agents/runtime-temporal, @helix-agents/runtime-dbos and @helix-agents/agent-server.

endStream requires the run's status

// Before
endStream(streamId: string, finalOutput?: unknown, terminal?: {...}, fence?: StreamRunFence): Promise<void>;

// After
endStream(
  streamId: string,
  finalOutput: unknown,
  terminal: StreamEndTerminal, // { status: 'completed' | 'interrupted'; state?; usage? }
  fence?: StreamRunFence
): Promise<void>;

• If you call endStream yourself, pass the ended run's own status: endStream(streamId, output, { status: 'completed' }), or { status: 'interrupted' } for a stopped run. A run that failed is never ended: call failStream with its error detail.
• If you implement a custom StreamManager, refuse an end with no valid status (assertStreamEndTerminal from @helix-agents/core throws a typed validation_error), persist terminal.status as StreamMetadata.endStatus on every end, and return it from getStreamInfo. The store contract (@helix-agents/core/testing) checks both (ES.end-status-required, ES.end-status-recorded). An untyped caller that omits the status now gets that validation_error from every built-in manager instead of a stream with no recorded status.
• state and usage stay optional: a caller attaches them only when its own terminal persist committed.

Readers never guess completed

Core's resumable stream handler (the wire v2 terminal frame), AgentServer.streamAgent, the Cloudflare Durable Object's own /sse frame and @helix-agents/store-cloudflare's stream Durable Object read the recorded status. A stream that ended with no recorded status is a contract violation. It is answered as a retryable transport_stream_drop (core.streamDropTerminal) and logged at error as core.stream.end_status_missing ({ streamId }), so a StreamManager that drops the status shows up in your logs instead of reporting success. createResumableStreamHandler takes an optional logger for that line.

Behaviour changes by runtime

Where	Before	After
Temporal, CF Workflows: an interrupted run's stream	no status (read as completed)	interrupted
DBOS: an interrupted run's stream (settle, entry restore, workflow)	no status (read as completed)	interrupted
DBOS settleCancelledRun(…, 'failed')	ended cleanly (read as completed)	failed with a framework_cancelled error detail
runtime-js: a completion whose terminal persist did not commit	no status (read as completed by luck)	completed, with no state / usage
Core stop / reconcile / revived-stream ends (every runtime)	no status	the run's own status
CF DO /sse late joiner, store-cloudflare stream DO /end, /info	a missing status was written / read as completed	/end without a valid status is refused 400 validation_error; a missing one reads as none

The Cloudflare DO's reconcile wrapper keeps the caller's status and still adds the DO's committed state and usage to a status-only end, so a wake-finished run's end frame carries the same payload as a driven run's.