Add golden-batch and mid-course-correction figures for the batch chapter

[kgdunn]

Ten PNGs and their three generator scripts, in the house matplotlib dialect (Okabe-Ito palette, DPI 300, save() helper, output directory from argv), for the batch process monitoring and control chapter.

Companion PRs: the chapter that embeds these figures is kgdunn/pid-book#272; the modelling and control code they exercise is kgdunn/process-improve#541 (process-improve[control]>=1.80, which builds the mid-course prediction interval at the decision point). Merge order: process-improve first, then this PR, then the book PR (its PDF step stays red until these files are on main).

batch/golden-batch-figures.py (4 PNGs)

• golden-batch-nylon-spread.png: the bundled nylon batches; setpoint replication at sub-percent spread while the response spread is roughly 14 times wider.
• golden-batch-replay-spread.png: 200 simulator batches replaying the identical golden schedule (the same campaign the chapter regresses on Z); final titers 4.24 to 9.83 g/L (14.4% CV) around the 8.01 g/L reference.
• golden-batch-z-scores.png: the initial-conditions score plot of the chapter's own 200-batch training campaign with the A/B/C feed classes (81, 71, 48 batches) as overlapping ranges along the first component; nearest-centroid assignment agrees with the true class 84.5% of the time.
• golden-batch-variance-decomposition.png: the four variance shares (initial conditions 31%, within-batch disturbances 27%, noise 0.03%, interaction 41%).

batch/mcc-decision-point-row.py (1 PNG, schematic)

• mcc-decision-point-row.png: the unfolded row (11 pre-batch values, then 20 samples of each of 5 tags: 111 columns) drawn twice: a complete training batch, and a running batch at the day-4 decision point with its three kinds of column (known; missing and estimated; the schedule under consideration).

batch/midcourse-correction-figures.py (5 PNGs)

Every quantitative claim is executed on the simulator (corrected schedules re-simulated under the same disturbance seed), replicating evaluate_control_policies(y_target=8.0, random_state=0) at its new default dead band of 1.0 (the whole prediction interval must fall short of the target):

• mcc-monitoring-funnel.png: one poor batch's predicted final titer at every decision point with the 95% prediction band built at that decision point; it narrows from ±4.0 g/L at day 0.5 to ±1.75 at day 4 and ±0.9 at day 9.5.
• mcc-correction-at-k.png: that batch's temperature and pH schedules before and after the day-4 correction (3.66 to 5.79 g/L executed).
• mcc-policy-comparison.png: 40 batches under four executed policies: replay 7.51 ± 1.20, mid-course 7.75 ± 0.78 (5 corrected, none harmed), oracle-from-k 7.87 ± 0.63, adapted feedforward 7.82 ± 1.01 g/L.
• mcc-decision-point-window.png: mean executed gain versus decision-point placement; the window is days 3 to 5 with the largest gain at day 4, and the early side no longer harms because the interval at the decision point refuses the ill-conditioned early projections.
• mcc-exploration-dial.png: predicted versus executed titer of the corrected batches as the T2 penalty is relaxed, caps off (executed 6.30 to 6.98 g/L, above the prediction at every setting).

Both data scripts print the numbers each figure rests on when they run. Regeneration commands are in each module docstring; the MCC set takes roughly 20 minutes because the policy comparison and the two sweeps are re-executed, not read from a cache.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NZAyiPT8xURSdi24yFKCra