feat(sdk): pod-runtime L5 — multi-agent collaboration primitives

[lIang70]

Summary

Lands the four high-priority L5 items from docs/sdk-pod-runtime-gaps.md
that close the remaining surface gaps in the multi-agent collaboration
layer. Each item was scoped down from the original RFC during review to
keep the change set incremental and avoid over-design.

kanban: Call(ctx, opts) synchronous sugar

• Subscribes to the board before Submit so a fast executor cannot
  race past the watcher and miss the terminal transition.
• Returns CallResult{CardID, Status, Output, Error, ElapsedMs};
  CardDone → nil, CardFailed → errdefs.Internal,
  CardCancelled → errdefs.Aborted.
• ctx cancel triggers a best-effort Cancel via
  context.WithoutCancel and returns ctx.Err(), preserving standard
  context semantics for callers.

kanban: Cancel + CardCancelled terminal status

• New terminal status distinct from CardFailed so failure-rate
  dashboards stay honest when a Pod controller stops in-flight work.
• Board.Cancel(id, reason) state machine: Pending/Claimed → Cancelled only; emits EventTaskCancelled + TaskCancelledPayload
  on Bus().
• Kanban.Cancel(ctx, id, reason) wrapper with NotFound / Conflict
  classifications and a new kanban.tasks.cancelled.total counter.
• Eviction, callback rendering, and subject helpers updated to treat
  Cancelled as a first-class terminal alongside Done / Failed.

history: LoadOptions filter

• New LoadOptions{Budget, Roles, SinceSeq, LimitN, IncludeTools}.
  SinceSeq is a position cutoff because model.Message carries no
  wall-clock timestamp — callers needing a time anchor look up the
  sequence index in their own audit log first.
• Optional FilterableHistory sub-interface keeps the History
  interface non-breaking; LoadFiltered() helper picks the most
  efficient path automatically.
• ApplyLoadOptions in-memory pass strips RoleTool messages and
  PartToolCall / PartToolResult parts when IncludeTools=false
  (the default for moderation / inspection use cases). Pure tool-call
  assistant turns become empty after stripping and are dropped.
• buffer + compactor both implement FilterableHistory via the
  shared helper.

graph/runner: stream.delta SPI promotion

• engine.EmitStreamToken / EmitStreamToolCall /
  EmitStreamToolResult / EmitStreamDelta package the
  SubjectStreamDelta envelope construction + header stamping +
  per-Type field validation so any node — not just LLM nodes — can
  publish in-flight increments without re-implementing the
  boilerplate.
• Malformed deltas are caught at emit time instead of silently flowing
  to subscribers.
• engine and graph/runner package docs explain the two paths
  (graph.StreamPublisher.Emit vs direct EmitStream*) and that both
  land on the same SubjectStreamDelta channel.

Test plan

• make vet — sdk, sdkx, voice, bench, examples, tools/conformance
• make test — sdk, sdkx, voice, bench, examples, tools/conformance
• 18 new test cases:
  • sdk/kanban/call_cancel_test.go: Call happy path / failed /
    ctx-cancel; Cancel state machine; Board.Cancel event publish.
  • sdk/history/load_options_test.go: zero-opts no-op, role filter,
    SinceSeq cutoff, LimitN tail-bias, tool stripping behaviour,
    IncludeTools=true, FilterableHistory fallback.
  • sdk/engine/stream_emit_test.go: token happy path,
    tool_call/result required-field validation, nil-publisher no-op,
    empty-Type rejection, forward-compatible Type acceptance.

Made with Cursor