examples/static_net: slots instead of Features/Slot<i>/Store<i>; compare_emlearn and the generators with it

[neu-rah]

What

examples/static_net on the slot core merged in #5 (hapi/slots.h). The values a net reads are now the slots of a state, read by tag: Field<Tag,&S::m> for a field, Elem<Tag,&S::arr,k> for an element of an array field. No hand-typed slot index is left (the old Features<N> array and its Slot<i> / Store<i> / x<i> / Wave<n> / In<n> / T<n> are gone): inputs are named (the Banknote features by their dataset names, a 60-band input is one array field), cells stay pure proc(in), and a value that survives a pass is a register, a layer of the state whose step evaluates cells on the previous state (include/registers.h, the example's own contract on the core; update/Store<i> wrote into the array being read, which cannot be right for two registers that read each other).

The example uses the HAPI core alone: no names, no hash, no wire format, nothing from another library.

In the same change, so the example is never half-converted:

• compare_emlearn/: gen_models.py names the four features once and writes BNC_COL_* and typed sources; the committed bnc_rows.h, bnc_lin.h, bnc_wave.h, bnc.cpp are what it writes. bnc_table.h and the emlearn models stay positional on purpose (they are the hand-indexed comparison points).
• train/bn2x.c and gen_lin_sonar.py emit the named forms directly, so a retraining cannot reintroduce a positional mapping.
• README (this example's and HAPI's paragraph), CHANGELOG (the 0.8.0 bullet).

Checked, against the example as it is on main (18b8d77, extracted with git archive)

• Same behaviour: the 12 host programs of check/ print the same bytes (g++ and clang++, ASan+UBSan); src/main.cpp prints agree 274/274, correct 274/274.
• Same code: the flashed image (.text + .data) of 21 AVR programs is byte-identical (27 372 bytes: mixed, refid, sugar, the Sonar cells, roll unrolled/rolled/sparse, the Banknote board program, the simavr bit-exact program, the measure/ programs).
• check/build.sh: 65 ok, 0 FAIL (the 61 checks it had, sizes unchanged, plus 4 for registers). tests/single_header/run.sh: 36 ok (the converted example's AVR programs and the bnc models are identical built from include/ and from single/hapi.h).
• compare_emlearn: run.sh over every model on the simulated ATmega328p regenerates results.md identical to the committed one; gen_models.py, run (numpy 2.5.3, scikit-learn 1.9.1, emlearn 0.23.2), writes all 22 generated files exactly as committed; the Sonar pipeline and the 8-minute bn2x retraining reproduce models/sonar/ and models/banknote/ byte for byte.
• One register example is wrong in the old design and right in the new: (a,b)' = (b, a+b) in place was wrong on 30 of 30 passes, on the previous state 0 of 30 (check/registers_check.cpp; the same step as a hand-indexed array is the same size, registers_avr.cpp).

Not checked

platformio.ini's own environments were not run (the file://../.. dependency copies a whole tree); src/main.cpp was built with the same flags natively and compiled for the Uno against the Arduino core (syntax only). CI does not build this example (GCC/Clang, AVR), so its checks are the ones listed above.

🤖 Generated with Claude Code