Add waterfall-progression regression test (incremental bucket flags)

[briantobin-99]

cc:
Closes:

What changed? Why?

Adds an integration test guaranteeing the release-waterfall progression keeps working, and fixes a B year-scaling regression on the legacy-footing configs.

Test. Toggling the bucket flags incrementally (v0 baseline → use_cornerstone_ghg_model → + implement_waste_disaggregation → + apply_io_year_adjustments + margins → usa_ghg_data_year: 2024, via the v03_waterfall_* configs) must reproduce each step's q-weighted average total emission factor — the release waterfall convention:

wavg N = Σᵢ Nᵢ qᵢ / Σᵢ qᵢ, with N = 1ᵀ B L (total-requirements L) and each step weighted by its own run's gross-output vector q.

• New bedrock/utils/validation/waterfall_progression.py — computes wavg N for one config, deriving N exactly as the EF diagnostics derive N_new, with a CLI and a --v0-baseline mode recomputing the baseline from the frozen v0 snapshots.
• New test_waterfall_progression.py (eeio_integration-marked, so it runs on the twice-weekday scheduled integration workflow) — runs each step in a fresh subprocess so functools.cache state cannot leak between configs, and asserts each step's level against pinned expected values whose per-sector N vectors match the N_and_diffs tabs of the pinned diagnostics sheets in release_v0_v03_ceda_groups (the sheets behind the v0.3 assessment) to ~5e-10 across all 405 sectors. Tolerance is 1e-4 kgCO2e/USD against step bars of ~1e-2, and v03_waterfall_final must telescope to the last step.

Fix. derive_cornerstone_B_non_finetuned keys the legacy-footing B year treatment on the bucket vocabulary: when use_ghg_year_x_in_B is false, B is scaled to usa_io_data_year (scale_cornerstone_B) and inflated to model_base_year (inflate_cornerstone_B_matrix_with_industry_pi); when true, B stays plain-vnorm. This restores the pre-#482 semantics for legacy-footing configs, which #482 had inadvertently dropped by re-keying the treatment onto the experimental use_scaled_x_and_scaled_Vnorm_for_B flag. The v0.3 path is unaffected.

Testing

All six waterfall tests pass, and running with USA_CONFIG_FILE exported into the environment (the scheduled-workflow condition, where earlier suite tests in the same process set it) produces identical values. Unit suite, black, ruff, and mypy green.

[bl-young]

I like this idea (sorry I know its still in draft), but would it make sense to build this test around the weighted average N waterfalls instead of the BLy waterfalls that we shifted away from?