Exercise 1: Smaller Checkpoints — Possible Solution
====================================================================

RESULTS ACROSS SEVERAL CHECKPOINT SIZES
------------------------------
    checkpoint_size=1:  1/20 rework (5%)
    checkpoint_size=2:  2/20 rework (10%)
    checkpoint_size=5:  5/20 rework (25%)  <- chapter's own baseline
    checkpoint_size=10: 10/20 rework (50%)
    checkpoint_size=20: 20/20 rework (100%)

At checkpoint_size=2 specifically: 2/20 rework (10%), an improvement
over the chapter's own 25% at checkpoint_size=5.

DOES IT KEEP IMPROVING, OR HIT A FLOOR?
------------------------------
The relationship is exactly linear across this whole range:
rework percentage = checkpoint_size / N_FEATURES * 100. There is no
floor within the tested range - smaller checkpoints keep producing
proportionally better results all the way down to checkpoint_size=1
(5% rework, the theoretical minimum for this scenario, since at least
the very first feature has to be built before any feedback is
possible at all).

WHY THIS WORKS AS AN ANSWER
------------------------------
This is a genuinely different shape than Chapter 1's own interruption
simulation, which was a sharp step function (all-or-nothing depending
on whether capacity was exceeded). Here, rework scales smoothly and
proportionally with checkpoint size - there's no threshold effect,
because unlike the interruption scenario (where switches either happen
or don't), this scenario's cost is a direct, linear function of "how
much got built before anyone looked." The practical implication: for
this specific class of risk (a wrong shared assumption), there's no
diminishing-returns point where shrinking checkpoints further stops
helping - the only real-world limit is the practical cost of gathering
feedback that often (a genuine tradeoff this exercise's own numbers
don't capture, but Chapter 1's own too-much-process findings do).