test: give the debug cases and annual integration their own predbat instance

[springfall2008]

Fixes the order-sensitivity that has already forced two golden regenerations and produced several confusing failures.

The problem

run_debug_cases planned against the shared my_predbat, and read_debug_yaml only restores the attributes its dump actually carries — so anything the dump omits inherited whatever the previous test left behind.

Two consequences, both real:

• The goldens depended on test ordering. Reordering TEST_REGISTRY changed which state leaked in, which changed the plan, which broke predbat_debug_pre_saving1 and forced a regeneration that had nothing to do with the planner.
• The suite and the CLI disagreed. --debug cases/<case>.yaml standalone produced a different plan from the same case run inside the suite, so reproducing a suite failure by hand didn't reproduce the plan.

test_single_debug.py already carried a hand-maintained list of attributes to reset for exactly this reason — dynamic_load_baseline, battery_rate_max_export, rate_max_base, rate_export_max_forward. That is whack-a-mole; the next omitted attribute starts it again.

The fix

Each case gets a freshly created instance, built the same way the standalone path builds one.

Verified predbat_debug_pre_saving1 now produces a byte-identical plan in all three contexts:

• --debug cases/predbat_debug_pre_saving1.yaml standalone
• ./run_all --test debug_cases
• the full ./run_all

Previously the first differed from the last.

What that uncovered

Removing the leak removed a mask. debug_cases had been overwriting most of the shared instance immediately before annual_integration ran, and annual_integration never sets up the state it plans against. Left with the ambient state instead, it went from 34s to 463s — silently, still passing.

It gets a fresh instance for the same reason. That also makes it faster than it was before any of this: 55s → 34s, because it is no longer planning against whatever a previous test happened to leave.

Effect

	before	after
full ./run_all	144.77s	119.28s
annual_integration	55s	34.10s
debug_cases	~25s	24.76s

Faster, and no longer dependent on what ran before what.

./run_all and ./run_all --quick both pass; pre-commit clean.

Note

The per-attribute resets in test_single_debug.py are now redundant, since a fresh instance starts from the same defaults they were restoring. I have left them: they are harmless, and removing them is a separate change that wants its own verification.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR removes order-sensitive state leakage in the unit-test harness by ensuring tests that load partial debug dumps (and tests that implicitly depend on prior state) run against freshly created PredBat instances, matching the standalone --debug <case> execution path.

Changes:

• Update run_debug_cases to create a new PredBat instance per debug YAML case instead of reusing the shared my_predbat.
• Introduce run_annual_integration_isolated and route annual_integration through it so the integration test runs against a fresh instance rather than inherited ambient state.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.