[subidhkhanal]

Fixes #375

Problem. A decision step auto-holds (no trade, the agent's decision skipped) at one of four places when its deadline has passed. #374 added a printed warning at three of them — get_current_step, drain_expired, get_status — all of which go through the shared _maybe_apply_timeout. submit_decisions had its own, independent deadline check that bypassed that helper entirely, so a hold created by a late POST printed nothing. Two real routes reach this: the legacy /api/v1/backtest/* surface (Discord bot, built-in agents) and the canonical /api/v2 contract. Both are ordinary GET-then-POST flows where the deadline can elapse in the gap, which is exactly the case this feature exists to surface. The periodic reaper doesn't catch it either, since submit_decisions has already advanced the step by the time it runs.

Fix. submit_decisions's late-submit branch now goes through the same shared _maybe_apply_timeout the other three sites use, instead of its own inline check, and prints the warning via the existing _emit_deadline_hold_notice helper on a hold. Routing through the shared function means there is only one place in the code that decides "this step is late" — a future change to that rule can't miss this spot the way #374 did.

Tests. Two new cases in test_deadline_hold_visibility.py:

• a late submit prints exactly one warning line, naming the held step;
• an on-time submit prints nothing.

I verified the new late-submit test actually catches the bug: run against the pre-fix code it fails (0 warnings printed instead of 1).

Verification run locally:

pytest dashboard/backend/tests/test_deadline_and_holds.py dashboard/backend/tests/test_deadline_hold_visibility.py dashboard/backend/tests/test_external_backtest_api.py dashboard/backend/tests/test_external_minute_source.py dashboard/backend/tests/test_external_session_compat.py dashboard/backend/tests/test_protocol_api.py dashboard/backend/tests/test_v2_runs.py -q

84 passed. Two pre-existing failures in that run also fail on unmodified main on Windows and are unrelated (Windows clock-resolution flakiness in deadline-timing tests — real-clock comparisons at a 0-second deadline can land exactly on the boundary; both pass reliably on Linux/CI).

Review follow-up (dd54d9f). Both items from the scope note are fixed here, plus three related ones from review:

• One helper, _apply_expired_holds, now handles hold-and-print for all four sites. It is the only caller of _maybe_apply_timeout.
• A late hold of the final step returns run_id/metrics/compare_url. On v2 it also marks the protocol_runs row completed right away.
• A decision built for a step that already closed is refused (409 step_already_closed) instead of being applied to the next step. That covers the reaper or a status poll holding the step first, and a hold landing between v2's step read and apply. Mechanism: an expected_step pin plus served-step inference.
• v2 audits idempotency and trace records under the step that actually executed or was refused.
• The accepted ack reports decision_source again, instead of None.

🤖 Generated with Claude Code

[vercel]

@subidhkhanal is attempting to deploy a commit to the allan-feng's projects Team on Vercel.

A member of the Team first needs to authorize it.