Clean up what the last audit turned over: two switches, six stale commands, and four crowded figures

[qmohsu]

Follow-up to #93. Three problems that surfaced while doing that work, each
fixed and each guarded where a guard applies.

1. Two mechanisms decided whether a window opens

#80 made that one decision: core.eval.show_figures_if_requested reads
IPIN_SHOW_FIGURES and is off by default. Two examples kept a --no-show flag
from before it, gating the new helper:

if not args.no_show:
    show_figures_if_requested()
else:
    plt.close("all")

So the reader had two switches for one behaviour, and the flag only did
anything if they set the environment variable and then asked for the window
to be suppressed anyway. Exactly the shape #80 removed everywhere else, left in
the two files it did not touch. Gone from
ch2_coords.example_attitude_visualization and
ch7_slam.example_scan_matching_visualization, with the branch, the two tests
that passed it, and the ch7 README line.

2. A module docstring is reader documentation now — six commands had not caught up

Fixing ch2's docstring, I said it was the only one. git grep said otherwise:
six more, all in Chapter 3.

ch3_estimators/example_comparison.py
ch3_estimators/example_ekf_range_bearing.py    (2)
ch3_estimators/example_iekf_range_bearing.py
ch3_estimators/example_kalman_1d.py
ch3_estimators/example_least_squares.py

python ch3_estimators/example_least_squares.py is the invocation that fails
with ModuleNotFoundError on a clone that has not been pip-installed — the
whole reason the repo standardised on python -m. They survived because
test_documented_commands_use_module_form read .md and .ipynb only.

Excluding .py was reasonable before #80 and is not now. Every example
passes description=__doc__ to argparse, so a module docstring is printed
verbatim by --help. The guard reads .py too; three files that quote the
script form as the failing case are exempt with that stated. Mutation-tested.

3. Chapter 2's figures were 85% scaffolding

The four figures #93 surfaced are good, and matplotlib's default 3-D axes
surround each unit triad with three filled grey panes and a full grid that take
most of the area and none of the attention. There are no numbers to read off
these axes — the panels say "this arrow moved, that one did not".

Panes unfilled, edges and axis lines at 0.75–0.85 grey, grid at alpha 0.15,
tick marks removed. One shared _style_3d helper, so all four improve at once,
and the triads are now what you see. Each was regenerated and opened before
being committed.

ch8's nine-panel lc_tc_comparison was considered and deliberately left: it
appears a third of the way down, not on the first screen, and it is the
chapter's core comparison rather than an attention-grabber. Redesigning it is a
different change with a much less clear payoff.

Verification

Full suite running locally; every README transcript and every committed figure
gate must hold, and the twelve regenerated Chapter 2 files are the only figure
bytes that move.

🤖 Generated with Claude Code