fix: finalize circuit-broken agent runs through usage accounting

[orange-dot]

Summary

This fixes the circuit-break finalization path so agent runs record and publish usage through the same run-level accounting surface used by the normal completion path.

Why

Circuit-broken runs can terminate after real model usage has already accumulated. If that path does not finalize through the same run-level accounting source as normal completion, the recorded totals can drift from the actual run state.

What changed

• read usage from agent_run.usage() in the web agent flow
• use that run-level usage in the circuit-break finalization path
• align the normal completion path to the same source for consistency
• keep the supporting loop-guard and discovery-tool behavior changes in the same PR because they belong to the same runtime-correctness surface
• add focused regression coverage for the circuit-break accounting path

Scope guardrails

• no schema or API changes
• no protocol or wire-format changes
• limited to agent runtime finalization/accounting behavior and the directly related guard behavior needed to exercise it cleanly

Validation

• isolated agents_test DB recreate plus agent migrations
• docker compose -f docker-compose.dev.yml run --rm -e DB_NAME=agents_test agent uv run --group dev python -m pytest tests/test_agent_http.py tests/test_tool_loop_guard.py tests/test_list_integration_resources.py
• docker compose -f docker-compose.dev.yml run --rm -e DB_NAME=agents_test agent uv run --group dev python -m pytest tests

Observed results:

• focused regression suite passed: 12 passed
• full agent suite passed: 172 passed
• the signed-off single-commit snapshot was rerun as-is after metadata cleanup

Request for review

Would appreciate a close look at whether this keeps the accounting/finalization contract tight enough around circuit-broken runs without widening scope beyond the runtime paths that actually participate in usage publication.

[superplanehq-integration]

👋 Commands for maintainers:

• /sp start - Start an ephemeral machine (takes ~30s)
• /sp stop - Stop a running machine (auto-executed on pr close)

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4d3be70. Configure here.}

[orange-dot]

Observation from local follow-up on April 16,
2026:

The public Semaphore run failed on
TestSendEmailComponent/ running_send_email_in_a_canvas_flow with timed out waiting for execution of node Send Email.

I ran the same targeted test locally in an
isolated compose setup against:

• PR head
  fb75cd0e710847523b406cf3dfe274e08eb645be
• PR merge ref
  e734a2d5724d0b1ea9fd212692b7437c45df59ab
• local main baseline
  62671f252985284d6d468e07a10b58d05fe5fd86

All three passed locally.

Not claiming the PR is clean from this alone,
only noting that I could not reproduce the
current Semaphore Send Email failure locally
on either the head or merge snapshot.