mattandClaude Sonnet 5
afa02864a7
CI / test (push) Successful in 12s
Two tests were spinning up real CBC solves over 96-half-hour windows with long runs of exactly-repeated prices (e.g. [0.0] * 40). That gives the LP relaxation a huge set of economically indistinguishable ways to spread a trade, which the MILP fallback's branch-and-bound then wastes enormous effort disambiguating (47k+ nodes without closing the gap, confirmed by running CBC verbosely). Real Attachment 2 data has no such flat runs and solves in ~0.1s/window; a synthetic sine wiggle wasn't enough either, since neighbouring half-hours stayed too similar. Fixes, matched to what each test actually needs: - The 9 validator tests only need *some* structurally valid schedule to mutate; they were deriving it by running the real optimiser once per test. Replaced with make_valid_schedule(), built directly from the model's own energy-balance formulas -- no solver involved, and it's now a true unit test of validate_schedule() in isolation. - The rolling-horizon carry-forward test was exercising the mechanism at full production scale (48h window / 24h commit) when a 2h/1h window proves the same boundary-carrying behaviour with a trivial MILP, regardless of price structure. - Fixed a genuine tie in test_optimum_uses_whichever_market_pays_more: two equal-price hours with just enough stored energy for one meant either market was a valid optimum. Sized the charge phase so delivery must split across both hours, pinning a unique answer. Full suite: 38 passed in ~1.3s (previously hung indefinitely on CI and locally within seconds of the same wall-clock variance CBC shows on degenerate MIPs). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>