Fix seeding bugs and correct release metadata for 0.3.0

[jshaw]

Code review of the tagged 0.2.1 release. arduino-lint --compliance strict passed clean on 0.2.1 and all examples compiled, so the problems here are not spec violations — they are metadata that is valid but wrong in effect, plus four correctness bugs.

Ships as 0.3.0, not a patch: the PRNG change means a given seed produces different noise than it did in 0.2.x. Anyone who saved a seed to reproduce a pattern will need to re-pick it.

Correctness fixes

reseed() before the first noise() call was silently discarded. It regenerated the table but never set initialized, so the lazy init in noise() immediately overwrote it with init(millis()).

Seed 0 did not seed, and the automatic seed rarely varied. Seeding went through randomSeed(), whose AVR implementation is if (seed != 0) srandom(seed); — so init(0)/reseed(0) were no-ops. init() with no argument passes millis(), which is ~0 that early in a sketch, so the "automatic random seed" path frequently produced an identical permutation on every board and every boot. Replaced with a private xorshift32, which also makes a given seed reproducible across AVR/ESP32/Teensy (Arduino's PRNG is not the same on all three) and stops generatePerm() from clobbering the sketch's own random() state.

fbm() with octaves < 1 returned NaN. maxValue stayed at 0.0, and scaledFbm passed the NaN straight into servo.write(). Now clamped to at least 1.

generatePerm() also now shuffles in place within perm[0..255] instead of into a separate 256-byte stack buffer that was allocated lazily from inside noise() — on an Uno the servo example leaves 584 bytes for locals, so an unpredictable 256-byte frame was worth removing. This fell out of the rewrite rather than being separate work.

Packaging fixes

Dropped depends=Servo,ESP32Servo,PWMServo. library.properties has no per-architecture dependency mechanism, so this installed all three servo libraries on every board — for a library whose core noise API needs none of them. I confirmed the Teensy consequence locally: with Library Manager's PWMServo installed alongside the Teensy core's bundled copy, the resolver reports Alternatives for PWMServo.h: [PWMServo@2.1 PWMServo@2.1] and picks the sketchbook copy over the core's. The README now documents the per-board install as a table instead.

includes=SimplexNoise.h — previously also listed SimplexServo.h, so the IDE dragged a servo library into noise-only sketches.

License. The LICENSE file was GPL-3.0 while the README and both source headers said public domain. GPL-3 on an Arduino library is strong copyleft that propagates into users' sketches, which is almost certainly not the intent and contradicts Gustavson's public-domain original. Now MIT, with license=MIT in library.properties and the upstream public-domain attribution preserved.

⚠️ Reviewer note: relicensing is effectively one-way. There is one merged outside contribution (szekelyisz, a24ea8e), though that code was fully rewritten in v0.2. Worth a second look before merging.

keywords.txt no longer advertises mapNoise, fastFloor, dot, F2 and G2 — all private.

Documentation fixes

The "What's New in v0.2.1" list claimed "Memory optimization: Uses PROGMEM for constant data to save RAM". Checking the v0.1 tag, static const uint8_t p[256] PROGMEM and pgm_read_byte were already there — that bullet was never new. Also replaced the "True randomness" feature claim, added Library Manager install instructions, and pointed the release checklist at arduino/library-registry for the first-time submission.

CI

Stopped installing PWMServo from Library Manager. The Teensy core bundles it, and the LM copy shadowed the bundled one — CI was validating something Teensy users don't actually get. Verified Teensy still compiles without it.

Verification

• arduino-lint --compliance strict --library-manager submit — clean, no errors or warnings
• 3 examples × arduino:avr:uno, esp32:esp32:esp32, teensy:avr:teensy41 — 9/9 compile
• Host harness with an Arduino shim: seed determinism, seed 0 honoured, distinct seeds diverge, fbm(0) and fbm(-3) finite, output ranges sane over 20k samples
• Bug A confirmed in two fresh processes — reseed(555)→noise() and init(555)→noise() now both yield 0.39794308497523956
• Shuffle validated across seeds 0, 1, 555, 12345, 0xFFFFFFFF: valid 256-permutation, upper half mirrored, permMod12 consistent

AVR cost: ~+142 bytes flash, +2 bytes RAM (xorshift is slightly larger than the random() calls it replaced).

Deliberately out of scope

• permMod12[512] is 512 bytes of pure cache for perm[i] % 12; with perm[512] that's 1024 bytes, and the noise-only example still sits at 64% of an Uno's RAM. Biggest remaining win.
• double throughout is free on AVR but software-emulated on ESP32 and Teensy 4.x. Switching to float would be a real speedup on two of the three CI targets but changes public signatures.
• Source header comments still say "(2025)" while LICENSE now says 2026.

🤖 Generated with Claude Code