fix: make two CI tests hermetic to the wall-clock window

[johanzander]

Summary

• Fix a Fast-tests flake: test_app_startup.py loads the real app.py, whose
  BESSController.__init__ calls set_timezone("UTC") from the mocked HA
  config — leaking the shared module-global time_utils.TIMEZONE into the
  rest of the test session. At UTC hour 0 the historical-data endpoint then
  computes zero "missing" hours, so test_dismiss_persists_across_requests
  always gets dismissed: False. Restore TIMEZONE on teardown.
• Fix an E2E flake: the vpp_power schedule spec mocks a 00:00–01:00
  period group, which the page treats as "current" whenever the real wall
  clock falls in that window — the Current Strategy card then renders a
  second "+35% (Remote)" and the strict getByText('+35%') resolves to 2
  elements. Scope the assertion to the schedule table's VPP cell.

Root cause

Both are the same class: tests not hermetic to the wall clock, plus CI runs
landing in the 00:00–01:00 UTC window. Neither is caused by #671 or #672 —
both PRs hit the identical failures on their own CI runs.

Verification

• pytest backend/tests/test_app_startup.py backend/tests/test_dashboard_api.py::TestHistoricalDataStatus
  failed before (1 failed) and passes after (8 passed).
• Full pytest -m "not slow": 2225 passed / 0 failed.
• E2E fix verified by locator analysis (single cell match); full E2E suite
  requires the mock-HA/podman stack and runs on CI.

🤖 Generated with Claude Code

[bess-agent]

@claude-bot review

[bess-manager-claude-bot]

Review: PR #673 — fix: make two CI tests hermetic to the wall-clock window

Verdict: APPROVE

No linked issue — this is a self-contained CI-flake fix (PR body references #671/#672 only to show the flake isn't caused by them).

Fix 1 — backend/tests/test_app_startup.py (+8/-0)

Root cause matches the diagnosis exactly. Verified in core/bess/time_utils.py:

def set_timezone(tz_name: str) -> None:
    global TIMEZONE
    TIMEZONE = ZoneInfo(tz_name)
...
TIMEZONE = ZoneInfo("Europe/Stockholm")

TIMEZONE is a genuine module-global mutated in place by set_timezone. _load_real_app_module execs the real app.py, whose BESSController.__init__ calls set_timezone("UTC") from the mocked HA config (get_ha_config → {"time_zone": "UTC"}), permanently flipping the shared global for the rest of the session — exactly the leak the PR describes.

The fix:

import core.bess.time_utils as time_utils
monkeypatch.setattr(time_utils, "TIMEZONE", time_utils.TIMEZONE, raising=False)

This is a correct, idiomatic use of monkeypatch: it snapshots the pre-exec value of TIMEZONE before exec_module runs set_timezone("UTC"), and pytest restores that snapshot at the function-scoped monkeypatch fixture's teardown — regardless of how many times the global gets reassigned during the test body. Since each test function in this file takes its own monkeypatch, the restore happens at the end of every test that calls _load_real_app_module, which stops the leak at the file boundary as claimed. No workaround shape here (no new flag/param/second construction site) — this fixes the actual leak at its source using the tool built for exactly this.

Minor nit: raising=False is redundant since TIMEZONE is unconditionally defined at module scope — harmless, not worth a changed request.

Fix 2 — e2e/tests/inverter-schedule-control-model.spec.ts (+5/-2)

Confirmed against frontend/src/components/InverterStatusDashboard.tsx: the "Current Strategy" card (line 637, a StatusCard) renders +35% (Remote) as a keyValue/metric string — not a table cell — whenever dashboardControlModel === 'vpp_power' and the wall clock falls inside the mocked 00:00–01:00 "current" period group. The schedule table's VPP column cell (line 916) independently renders +35% inside a <td>. The old page.getByText('+35%') (strict mode) would match both nodes whenever a CI run landed in that UTC hour — exactly the flake described. Scoping to page.getByRole('cell', { name: '+35%' }) correctly excludes the card (non-table markup) and, since the mocked periodGroups array has exactly one entry, uniquely matches the one table cell. This is a real fix, not a loosened assertion — it still asserts the signed-percent rendering, just via a selector immune to the day's actual wall-clock hour.

Checklist

• Matches root cause: yes, for both fixes — verified against the actual time_utils.py global and the actual card/table markup, not just taken on faith.
• Would the "test" catch a regression: yes — the existing tests these fixes touch (test_dismiss_persists_across_requests and the vpp_power E2E spec) already assert real behavior (dismissed-flag correctness under real "missing hours" math; correct VPP-column rendering). Nothing was weakened to make the flake go away.
• Minimal / no scope creep: yes — only the two files needed are touched, no production code changed, no unrelated cleanup.
• Rule violations: none found. No Optional[x], no hasattr/getattr fallbacks, no new classes, no ad-hoc API dicts, no string-matched exceptions, no hardcoded entity IDs — none of these apply since this is test-only. Comments added explain non-obvious WHY (the leak mechanism / card-vs-cell ambiguity), consistent with the comments rule.
• CI on this PR is fully green (Fast tests, E2E tests, Code quality, Docker build & boot), which is itself evidence the E2E fix works against the real rendered page, not just locator analysis.

One process nit (not a blocker): per rules.md Debugging Protocol step 9, the PR description doesn't explicitly label this as a "local fix" — but the diff itself makes that unambiguous (test-only, no new method responsibilities, no routing-around-a-problem shape), so I'm not treating the omission as a rule violation here.