Round 3 held-command continuation analysis

Entry point: analyze_baselines.py; helper: paired_helpers.py.
Dependencies: Python 3.12+, NumPy, pandas.

Example (from bundle root):
python sr_round3/analysis/analyze_baselines.py --input-dir sr_round3/baselines/results/archived_conditional --output-dir reproduced/archived_analysis --analysis-role archived_conditional
python sr_round3/analysis/analyze_baselines.py --input-dir sr_round3/baselines/results/causal_development --output-dir reproduced/causal_analysis --analysis-role causal_development

Confirmation is not automatically authorized by this file. It needs the
parent-approved pre-outcome freeze JSON, with confirmation_scoring_authorized
set to true, passed using --freeze-manifest and --analysis-role confirmation.
The producer's filename can be set using --filename.

Main quantities:
* primary paired gap = deployed shortfall frequency minus the observed failure
  frequency of the same-state/same-first-command frozen-forecast FULL backup.
* DMPC additional gap uses its own original optimized fixed command sequence.
  These are distinct estimands; their signs may differ without contradiction.
* F_FULL rolls forward with newly issued forecasts. It is not a zero-gap control.
  The genuine negative control replays the identical fixed command sequence.
* Target-matched Brier: backup scenario risk against backup labels; DMPC own
  scenario risk against own-plan labels. No rolling-deployment Brier is reported.
* All shortage and resource totals sum each actual executed row once; they do
  not sum overlapping horizon labels. Energy uses normalized-power hours.
* Complete-hour frequency requires six contiguous UTC steps, phases 0 through 5.

Bootstrap:
2,000 paired ordinary moving-block replicates, base seed 20261003; primary block
length 7 days, sensitivities 1 and 14 days. Blocks never cross missing-time gaps
or episode boundaries, and do not wrap. Each episode/run retains its original
sample count. A fragment shorter than or equal to the block length stays fixed;
reported fixed_fragment counts identify this conditional contribution. Ranges
exclude model-fitting, between-site and between-episode uncertainty.

The optional scoring_execution and scoring_origin masks permit continuous
warmup execution without including that warmup in the requested reporting
window. Without these arrays, all executed rows and valid origins are scored.
Each method must have exactly the same rows, masks, timestamps and horizons.

Validation stops analysis on missing risk targets, nonbinary labels, inconsistent
eligibility, unmatched first commands, ledger/label discrepancies, or solver
fallbacks. The physics simulator and input-time audit are reviewed separately.
