[burgulvedant]

Fixes #201

The baseline SchedulePolicy builds its lookup time as Timedelta(hod) + local_start_time.utcoffset(), where hod is the UTC hour of day. At UTC-7 the UTC hours 0 to 6 give a negative time. No schedule event starts before 0, so get_active_setpoint falls back to the last event of the day, the 19:00 night setpoint. For two hours every evening (17:00 to 19:00 Pacific) the schedule uses night setpoints while the comfort window is still active.

This adds a small helper, local_schedule_time, that wraps the sum into 0 to 24 hours, and uses it in SchedulePolicy._get_action_map.

I added schedule_policy_test.py with 8 tests: the wrap itself (negative offset, no wrap, local midnight, positive offset) and the day and night setpoint at UTC hours 0, 1, 2, 6, 7, 13 and 23 at UTC-7. Three of them fail without the wrap.

Not covered here: the offset is read once from the start time, so daylight saving changes during a run are not handled. That is a separate issue.

How I tested: I could not install the full TensorFlow 2.15 environment on my machine (Python 3.14), so I ran the new tests locally with the tf_agents and tensorflow imports stubbed out. They pass. The full pytest run on Python 3.10 and 3.11 will be done by CI.

• Tests pass (8 new tests pass locally with stubbed imports, full run pending CI)
• Appropriate changes to documentation are included in the PR (none needed, the new function has a docstring)

[burgulvedant]

CI note: the failing check on this PR does not come from the change. The test (ubuntu-latest, 3.10.15) job fails at setup because Python 3.10.15 is not available on the Ubuntu 26.04 runner that ubuntu-latest now points to. The other three matrix jobs were then cancelled by the matrix fail fast. No tests ran, so the 8 new tests in schedule_policy_test.py have not been checked by CI yet. Happy to rebase or adjust once the workflow is fixed.