[theAndreas]

The gamma correction is the exponential curve from the "Sparsamere Implementierung" section of Diskussion:LED-Fading. It stores 16 values for the top doubling and gets every lower doubling by shifting. The background is the article itself, LED-Fading.

What changed

• 0 maps to 0. The source's author points out that the exponential never reaches zero and recommends answering a setting of 0 with 0. That recommendation did not come over with the code, so 0 became 1. It is now answered as a separate case rather than by subtracting 1 from every result, because subtracting would also turn 255 into 254.
• Documentation. The header now explains what the curve is and links to both pages. It explains why 128 steps = 16 values × 8 bits, which a static_assert now enforces; before, numberOfBits() happened to return 8. It also notes what to know before changing the curve:
  • the coarse bottom end is a limit of 8-bit output, not of the code;
  • the dark middle (half the setting gives about 6 %) comes from the base of 256 that the shift trick fixes, while the article suggests a base of 10 to 100.
• Clean-up. Variable names now say what the values are (the old exponent holds the mantissa), and the empty sections are gone. memcpy_P stays, because that is how the firmware core reads flash everywhere.

What does not change

What reaches the strip stays the same today. DisplayColor::dimmColor deliberately keeps every channel at 1 or above, and a test enforces that, with or without the correction. A setting of 0 therefore still gives the faintest light. After this change, the correction passes on the same 0 that an uncorrected setting does, and the stage after it decides what that means.

Measured

• testGammaCorrectionCurve() checks both ends, that the curve never falls, and one value from each end of the shift range. Made to fail on purpose: against the old header, exactly the 0 check fails and every other check passes. The curve itself is unchanged.
• AVR flash: 48 116 → 48 140 bytes (+24), for the 0 check at its two call sites.
• Built: simulator with ctest, AVR, ESP32, RP2350, and both host-test suites. tools/documented-sizes.py passes for all three firmwares, and clang-tidy is clean on all four builds.

Whether the middle of the range should be brighter (a flatter base) can only be judged on the real strip, so this PR leaves it alone.