Second fixture: PJM 5-bus system with two arbitrary zones, solved over a long horizon

[micahpw]

Follow-on to #13 (single-week fixture makes monthly/annual gallery plots degenerate) and the annual-solve benchmarking: add a second, much smaller fixture so long-horizon views are cheap.

Shape:

• PJM 5-bus test system — Sienna's PowerSystemCaseBuilder.jl builds 5-bus variants directly from the package, same pattern as the RTS-GMLC fixture build
• Split the buses into two arbitrary zones/areas so area-level machinery (area dispatch, interchange, facets) has something to show
• Solve a long horizon (months → year): at 5-bus scale even a DCP annual solve should be CI-viable — the in-flight annual benchmark on RTS will calibrate expectations

Division of labor across fixtures (extends the #16 discussion):

• RTS-GMLC (1 week): nodal detail — line flows, reserves, unserved energy (Surface unserved energy (balance slack) as a first-class dataset #10), regression baselines
• PJM 5-bus (annual, 2 zones): time-dimension coverage — monthly/annual dispatch stacks (Gallery: single-week fixture makes monthly/annual plots degenerate #13), load duration curves, seasonal patterns

Open questions:

• Do the PSB 5-bus variants ship time series long enough for an annual solve, or do we attach synthetic annual profiles (scaled/repeated weeks)?
• One store with UC+ED like the RTS fixture, or ED-only to keep the annual solve fast?
• Cache/baseline mechanics: new fixture = new cache key + new regression baselines; follow the documented refresh procedure, iterate locally first (the gat-sienna image pattern generalizes — likely a SIENNA_SYSTEM build-arg rather than a second Dockerfile).

[micahpw]

Data point from the RTS annual benchmark (see #13): 365 daily UC+ED+Emulator steps on RTS-GMLC (73 buses) took ~4.7 h locally under DCP + slacks — ~46 s/step. At 5 buses with two zones, per-step cost should drop by well over an order of magnitude, putting an annual 5-bus solve plausibly under ~20–30 min — comfortably CI-viable and exactly the fixture this issue proposes. That makes #18 the confirmed vehicle for monthly/annual gallery coverage rather than any RTS horizon extension.

Practical note for whoever builds it: when iterating locally on macOS, bind-mount the output under $HOME (as the Makefile targets do) — Docker Desktop doesn't share /private/tmp, and the mount fails silently into the VM.