test: give each test its own PredBat rather than sharing one (#5079)

[chalfontchubby]

(Written by Claude, on behalf of Rik.)

Test infrastructure only - no user-facing or product code changed. One six-line change in apps/predbat/unit_test.py.

Problem

Every registered test runs against a single PredBat built once at startup. A test that mutates instance state and doesn't restore it therefore breaks whichever unrelated test happens to run next - depending only on the order they ran in. That's the pattern tracked in #5079, and it has produced a long series of confusing failures where the test that reports the error is not the test at fault.

Until now each one was fixed by hand-writing a snapshot/restore block naming the fields that particular test touched. That approach scales with the number of tests and has proven error-prone: the first attempt at one such fix missed 13 fields, and a follow-up audit found 7 more.

Change

Build the fixture per test instead, so leaked state cannot carry over at all.

Measurements

Known-bad orderings (the 53 (culprit, victim) pairs collected so far by the fuzzer in #5096):

	pairs failing
before	53 / 53
after	0 / 53

(The two involving plan_persistence need #5100 as well - that test also fails on its own, for an unrelated product bug.)

Runtime - the assumption that rebuilding would be too expensive turns out not to hold. create_predbat() is ~22ms, and the suite's cost is dominated by the prediction work inside the tests, not fixture setup:

	full suite	quick suite
shared fixture	174s, 316 passed	83s, 312 passed
per-test fixture	173s, 316 passed	80s, 312 passed

No measurable difference either way.

The tests themselves are sound: all 316 pass when run individually, each in its own process. So every ordering failure found was shared-fixture contamination, not a test quietly depending on an earlier one to set it up. (The single exception was plan_persistence, which failed alone because of the product bug fixed in #5100 - found precisely because isolation was checked.)

Consequences

• Supersedes the per-test restore patches in fix(tests): restore forecast_minutes/minutes_now leaked by two tests (#5079) #5097 and fix(tests): restore now_utc/midnight_utc leaked by test_octopus_url (#5079) #5098, which can be closed.
• Tests may now rely on a clean fixture rather than on whatever ran before them.

Test plan

• ./run_pre_commit passes
• Full suite (including slow tests): 316 passed, 0 failed
• All 53 known-bad orderings re-run individually: 0 failing
• All 316 tests run individually in isolation: all pass

[chalfontchubby]

CI red here is not this change - it is a pre-existing bug this branch is missing the fix for.

The run failed at 22:13 UTC, and plan_persistence Test 1 reported the plan it had just saved could not be loaded back. That is #5100: save_plan() computes its 8-hour expiry from self.now_utc (the fixture clock, pinned to midday) while storage.load() checks it against the real clock. Midday + 8h = 20:00, so from 20:00 UTC onward every saved plan is already expired when written. The same run on main at that hour would fail identically.

There is a second, opposite instance of the same mistake in Test 5, which builds an "already expired" timestamp the same way and so fails before about 11:00 local - fixed in #5104.

plan_persistence therefore only passes at any hour with both clock fixes present. Verified locally with all three stacked:

quick suite: 73s exit=0 passed=312 failed=0

Will restack this on top of #5100 and #5104 so CI here is meaningful.