Round 3 held-command controller reproduction

Requirements: Python 3.12, NumPy 2.3.5, SciPy 1.17.0. The LP uses
scipy.optimize.linprog(method='highs'); no CVXPY or separate highspy is needed.
Run from the Round 3 package's parent directory, or substitute absolute paths.

2021 causal development (all three frozen controllers, then matched risks):

python3 sr_round3/baselines/reproduce_baselines.py \
  --input sr_round3/causal_inputs/development/causal_development_2021.npz \
  --input-role causal_development \
  --input-manifest sr_round3/causal_inputs/development/input_manifest.json \
  --output reproduced/causal_development --workers 4

Historical-input conditional comparison:

python3 sr_round3/baselines/reproduce_baselines.py \
  --input sr_round2/inputs/calibration_2021.npz \
  --input-role archived_conditional \
  --output reproduced/archived_conditional --workers 4

External confirmation must use the input files created by the frozen input
pipeline and the authorized frozen cohort. It is not a model fitting command.
The confirmation manifest must identify role=confirmation,
split=confirmation_aemo_20260911_28, input_role=external_temporal_confirmation,
and the exact input_sha256. Only eligible timestamps in UTC
[2026-09-10 14:00, 2026-09-28 14:00) may be dispatched. Use:

python3 sr_round3/baselines/reproduce_baselines.py \
  --input PATH_TO_CONFIRMATION_INPUT.npz \
  --input-role external_temporal_confirmation \
  --input-manifest PATH_TO_CONFIRMATION_MANIFEST.json \
  --output reproduced/external_confirmation --workers 4

Output folders must be new/empty to preserve earlier evidence. Each method has
one NPZ, with the filename specified by result_filename in run_manifest.json.
The wrapper automatically appends separately labelled own-plan and common
backup scenario assessments. See output_schema.json for exact axes and Brier
target mappings. DMPC's original screen_risk concerns its OWN optimized plan,
not the common forecast-FULL backup. Do not mix these targets.

Actual contemporaneous wind is used only by the common local curtailment
actuator and retrospective scoring. All battery/reserve commands are fixed
before the interval. The horizon_mask and scoring_hour_complete arrays are
retrospective outcome metadata, never planner inputs. All three methods start
at E=.55, directional reserves .05 at each episode's first eligible row, even
when its hour is partial; incomplete hours are excluded from complete-hour
statistics. Source gaps require explicit new episodes and resource resets.

F_FULL and F_BR0 use the frozen origin point forecast for their continuation.
Future rolling forecasts are seen only when their origins are reached.
F_BR0 searches 225 first cap combinations; DMPC directly optimizes continuous
command trajectories under identical physical constraints. Their objectives
differ, so comparisons must include failure frequency, shortfall energy and
resource use. The proxy objective is not monetary operating cost.

Solver failures are recorded in full and use the explicitly specified F_FULL
fallback; invalid simultaneous actions or objective/actuator inconsistencies
stop execution. Report all solver statuses and fallback counts, including zero.
The controller core is hashed in code_manifest.json. A cohort-wide freeze of
input pipeline, frozen model, parameters and analysis belongs to the enclosing
Round 3 protocol, not this runner alone.
