feat(validate): CPUC Baseline Generator List capacity benchmark (CPUC stack 5/5)

[ktehranchi]

Stacks on #798 (base set to feat/servm-demand; retarget after the stack lands). Gated by a new run.benchmark_cpuc flag — useful for any California run, independent of the SERVM demand profile.

Changes proposed in this Pull Request

A validate-stage benchmark comparing installed capacity against the CPUC Baseline Generator List (June 2026), per planning horizon:

• workflow/scripts/benchmark_cpuc_baseline.py + rule benchmark_cpuc_baseline (runs on the assembled unsolved network — no solver needed). Outputs a long-format comparison CSV and a region × technology deviation heatmap PDF, surfaced through rule all when run.benchmark_cpuc: true.
• Region attribution: EIA publishes all CAISO plants as a single CISO BA (no PGE/SCE/SDGE split), so the benchmark compares CAISO-total (CPUC PGE+SCE+SDGE, which per the CPUC dictionary is the CAISO definition) plus individual non-CAISO regions (LDWP→LADWP, IID→IID, BANC+TIDC→NCNC), declared in repo_data/CPUC/servm_benchmark_regions.csv. Model side filtered to CA (EIA's CISO also spans the NV VEA footprint).
• Technology mapping: both sides map into common compare categories (repo_data/CPUC/servm_tech_map.csv) — SERVM's 27 tech categories are finer than PyPSA carriers. Unmapped techs on either side surface as explicit UNMAPPED: rows, never dropped; gas Cogen/CHP is kept CPUC-only (no CHP concept in the model — expect the model long in the gas buckets and short ~3.7 GW in Cogen, documented).
• Vintage filtering mirrors add_electricity.load_powerplants exactly on the model side — which makes visible (by design) that the model pins existing plants to a 2100 retirement and does not honor announced retirements. Note also a genuine boundary case in the CPUC data: both Diablo Canyon units carry RetireDates of 2029-10/2030-10 and therefore exit CPUC's own 2030 column.

Real-workbook sanity (2030, CA): CAISO 103.0 GW / LADWP 10.3 / NCNC 6.7 / IID 2.9; e.g. Battery 25.0 GW, Solar 30.8 GW, Gas CC 23.1 GW totals — table in the script's tests/docstring lineage.

Testing

• Tier A pytest -m fast: 195 passed, 1 pre-existing skip (combined stack).
• New tests: test_benchmark_cpuc_baseline.py — 22 passed (header layout, vintage boundary cases, unmapped-tech reporting, CAISO roll-up; plus a guard asserting the tech map covers every carrier in constants.EIA_TECH_MAP so new carriers can't silently vanish).
• DAG dry-runs: rule appears with the flag on (29 jobs), absent with it off (27), absent for default config; --until benchmark_cpuc_baseline pulls the workbook retrieve rule.
• Tier B not run for this PR. No equivalence run for this branch (flag defaults false; no existing outputs change).

Checklist

• I tested my contribution locally and it seems to work fine.
• pre-commit and pytest -m fast pass locally.
• Code and workflow changes are sufficiently documented.
• Changed dependencies are added to workflow/envs/environment.yaml (no dependency changes).
• Changes in configuration options are added in workflow/repo_data/config/config.default.yaml.
• Changes in configuration options are also documented in docs/source/configtables/*.csv.

🤖 Generated with Claude Code

[ktehranchi]

Superseded by #801 — the five-PR stack was collapsed into a single PR at the author's request. All commits (plus Tier B fixes) are on feat/cpuc-servm-california.