Recover the picture-spike measurements that never reached main

[mjaksn]

Summary

The picture-spike branch never merged, but most of what it measured on 2026-09-12 already reached main some other way, through PR #73's memory pool fix. This pulls out the small remainder that did not: findings that are still true and still useful, but sitting only on a branch nobody was going to look at again.

• docs/protocol-notes.md: settles one of the "three questions left open" after the first picture spike, how a picture is listed in the memory configuration (an ordinary eleven-character entry, type D).
• docs/protocol-notes.md: explains why a picture's own data is rows * columns / 2 bytes rather than rows * columns, a pixel takes four bits, packed two to a byte, which also confirms the 13-byte overhead nearby is real rather than a rounding artifact.
• docs/protocol-notes.md and readerboard/protocol/frames.py: corrects an overstated claim. Both currently say a read-back of the first file in a memory configuration "will therefore never match what was sent." Measured on 2026-09-12, that is only true once the sign has started playing a run sequence; a read taken right after configuring shows the plan as sent.
• A one-line historical note that an early draft icon set was tested directly on the sign before the built-in icon library shipped.

Test plan

• pytest
• ruff check .
• mypy readerboard