[KarthikAvinashFI]

Why

TH-10189 and TH-8888, plan tasks T16 and T17 ("Test a multi-agent application"): when the platform stopped a run mid-call, the hosted call workers never saw the stop, so the call in flight kept nothing; a separate worker that reached the world on the shared address was graded as if it had used its own channel; and nothing recorded when two workers touched the same record. The multi-agent fixture could also not fail on a duplicate refund, because its ledger refused one itself.

This PR seals each in-flight call as user_canceled with its partial evidence, flags a worker channel bypass and leaves that worker's checks unmeasured, records shared records, and makes the fixture able to fail. The platform side is future-agi feat/TH-10189-worker-outcomes.

What changed

File	Change
src/fi/alk/harness/hosted_scheduler.py, hosted_entrypoint.py	A user stop seals each in-flight call as user_canceled, pointing at its partial transcript; user_canceled is never retried; a runner that ignores the cancel is cut at the bound; a runner that fails while being stopped, or a call that fails under the stop, is sealed as the stop with what it kept, not as infrastructure. A scenario not yet started when the stop arrives returns user_canceled with no call. A bypass leaves that worker's checks unmeasured and still grades the task.
src/fi/alk/harness/isolated_process.py, call_worker.py, chat_worker.py	The stop reaches the chat and voice call worker subprocesses (SIGUSR1); the parent raises the stop with the worker's sealed call; a stopped worker exits once its result is written, not after a thread still blocked on the agent.
src/fi/alk/harness/call_runner.py	Voice: stop_when hangs up at the next turn. Shared records: records two or more participants named in a trial and what the world made of them (write from the contract's effects, unknown is null). Bypass: a separate worker's call on the shared address is flagged; shared tools, shared workers and door rows from before channels are not. A worker's door row and its framework span are one call. scenario_started is sent as the call starts.
src/fi/alk/harness/chat_call_runner.py, src/fi/simulate/agent/wrappers/http.py	A stopped chat call seals its partial transcript, and abandons the agent reply it was waiting on (a stopped run's world is never reused). A cancel that is not a stop still propagates.
src/fi/alk/harness/observability.py	Tracing keeps the harness's SIGTERM handler, so the platform's Stop seals the calls in flight instead of exiting the guest.
src/fi/alk/harness/world/runtime.py	Call.channel: the channel the world saw the call come in on (shared address, a worker's own channel, or none for rows from before channels).
src/fi/alk/harness/bundle_author_v2.py, generated_runtime.py	A Compose-selected Python agent is observed for its hand-offs and worker tool calls; a native contract whose tools call the tools-api service the source ships runs it behind the world's door; langgraph dev/up is served through its threads and runs API (assistant = the graph the written interface names), not a model-written chat route; framework instrumentors land in the agent's uv project when the command is rewritten to sh -c for the world's port.
tests/fixtures/harness_agents/returns_multi_agent/ (moved from examples/harness/returns_reference/), tests/fixtures/harness_agents/multi_agent_compatibility/ (moved), .github/workflows/sdk-smoke.yml	The multi-agent fixtures are internal CI fixtures, not examples. The returns ledger allows a duplicate refund, and gains fault switches (skip_refund_check, payment_times_out, refund_never_returns, bypass_channel), off by default. The compatibility references test runs in CI.
tests/harness/*	Tests below.

Behavior callouts

• A stopped call now produces a receipt with failure code user_canceled and its partial transcript, hand-offs, tool calls and worker record, where it produced nothing before. A platform without the companion PR does not read it as cancelled.
• A worker flagged for a channel bypass has its own checks reported unmeasured ("reached the world without their own channel ... their own checks were not graded"); the task is still graded.
• The returns fixture no longer refuses a second refund: a duplicate-refund scenario can now fail.
• examples/harness/returns_reference and examples/harness/multi_agent_compatibility moved under tests/fixtures/; nothing outside tests referenced them.
• No new env var in the harness. The fixture reads RETURNS_FAULT, RETURNS_HANG_SECONDS and RETURNS_TRACING (fixture only).

Tests

• 40 new tests, including test_cancel_mid_call_seals_a_user_canceled_receipt_pointing_at_the_partial_transcript, test_one_stop_seals_every_call_in_flight_each_with_its_own_transcript, test_a_runner_that_ignores_the_cancel_is_cut_at_the_bound, test_a_stopped_worker_exits_without_waiting_on_a_call_stuck_in_a_thread (60 s thread: exits in seconds, waited 60 s on the base), test_a_stopped_call_abandons_the_agent_reply_it_was_waiting_on (under 2 s, 5 s on the base), test_tracing_never_takes_over_the_harness_s_stop_signal, test_a_separate_worker_call_on_the_shared_address_is_a_bypass, test_shared_records_list_two_participants_on_one_order, test_the_ledger_creates_a_second_refund, test_langgraph_dev_is_served_through_its_threads_and_runs_api.
• Run module by module: 55 modules, one at a time. All new tests pass; the 32 failures that remain are identical on the base d46d591c. test_bundle_author_v2, test_authoring_runtime_validation and test_runtime_repair green (281, 109, 205). On this branch's final tree (without fix(harness): name a served coordinator's own calls by its run, as its jobs' hand-offs do #318), test_bundle_author_v2 and test_framework_tool_spans: 388 passed.
• Mutation checks: 29 killed (plan-named tests plus each fix test), plus 3 for the LangGraph serving.
• Break-it pass: one stop with two calls in flight, a stop landing as the call finishes (normal receipt), a stop before any call starts (every scenario skipped), a worker lost on the way out of a stop, a library replacing the guest's signal handlers.

Live proof

Local stack with production sandbox specs (4 CPU, 8192 MB, restricted egress) and the companion platform branch, on the returns fixture as a hosted agent.

Case	Result	Proof
Stop mid-call while the refund worker waits on payment (jobs 0cc6e62f, 2bde6f55, f66ff308)	Call cancelled, ended_reason user_canceled; transcript, hand-offs, tool calls and the workers record kept	Image 1
Stop before the call started (job 38c4153b)	Keeps nothing, no crash	Not shown
A build whose refund worker uses the shared address	"Refund worker reached the world without their own channel ... their own checks were not graded"	Image 2
Two workers touch order 1043	Listed as a shared record with both workers	Image 3
Payment call times out (payment_times_out)	issue_refund returns "payment service timed out"; the hand-off returns; no refund committed; the full-refund check fails as it should	Image 4
LangGraph async deep agents, hosted build	Before: the builder wired the graph server as an OpenAI chat route (404). After: build 11558728 completed; run c192e6ed shows the coder hand-off start and end with its result and run_code_check under the coder	Images 5, 6
Worker hangs past the chat timeout	Not reachable live: a new agent version is certified first, and a hang never passes certification. Unit tests only	Not shown

The live runs, including the LangGraph ones, used a local build that also had agent-learning-kit #318 applied (a served coordinator's own calls named by its run). #318 is a separate draft PR to dev touching other files; it is not part of this branch.

Playwright e2e

None in this repo. The platform participants flow (stopped call, cleanup, shared records) passed in the platform's flows/simulate 8 of 8 run with this branch mounted.

Scope out

• A refund hand-off still out at the Stop has no Participants row: its framework span is exported only when it ends.
• A per-worker stop: workers share the app's one process; only the whole call stops.
• LangGraph: launches read "parent unknown", and a researcher job that outlives its 30 s call is beyond the call's boundary (TH-9023).
• MCP, CLI and SDK; T44 evidence access.

Stacked on / merge order

• Base: feat(harness): worker permissions and deadlines at the world's door (TH-8887, TH-8888) #265 feat/TH-8887-worker-permissions-and-timeouts. Merge feat(harness): worker permissions and deadlines at the world's door (TH-8887, TH-8888) #265 first.
• Companion: future-agi feat/TH-10189-worker-outcomes reads the user_canceled seal, shared records and notice workers.
• Independent of fix(harness): name a served coordinator's own calls by its run, as its jobs' hand-offs do #318 (draft, to dev).

Linear: TH-10189, TH-8888