neural_tile: a self-assembling tile computer in the abstract tile assembly model. A tile binds at a site when the summed strength of its matching glues reaches tau, which is the Heaviside gate H(strength.match - tau), so growth is governed by threshold neurons. Verified: the binding decision equals the gate; a general 2-input rule-tile set grows value(x,y)=f(W,S) for f in XOR/AND/OR (529 tiles each, checked against the recurrence, XOR = Sierpinski/Rule 90); a binary counter grows one integer per row (8-bit, 255 rows, row y encodes y) with carry by cooperative binding; both directed (deterministic). Turing-universal at tau=2 (Winfree 1998). Ships variants/neural_tile.safetensors (glue tables + binding-gate weights); eval_all skips it; README section and counts updated (9 standalone machines, 28-file family).

neural_ca: the collision is verified as a reversible interaction gate, not just AND. At step 4 three output cells carry A&B (deflected, (3,6)/(6,3)), A&~B ((6,6)), and ~A&B ((3,3)), checked over all four inputs. Analog robustness sweep on the matrix tile: exact under read noise through sigma 0.10 and conductance mismatch through sigma_G 0.10, matching neural_matrix8. README updated.

neural_ca: construct and verify a logic gate in the automaton rather than citing one. A billiard-ball collision between input particles at (2,2) and (7,7) leaves the deflected cell (3,6) occupied at step 4 iff both inputs are present; and_gate() checks the AND truth table over all four inputs. README and docstrings restated as plain technical description of the rule, the verified dynamics, and the constructed gate.

neural_ca: a reversible cellular-automaton computer with no processor. One fixed Margolus block rule (rotate 180, diagonal pairs swap) applied identically to every 2x2 block of a lattice, alternating partition each step. The rule is a self-inverse permutation of the 16 block states, so the whole lattice update is a bijection; verified reversible over random lattices, particle-conserving, with ballistic single-particle motion and deflecting reversible collisions (the billiard-ball model, Turing-universal by Margolus 1984). The rule is Heaviside threshold gates and compiles to a 6-layer ternary matrix tile that is a permutation with a 0.5 margin; that one tile applied to every block is one whole-lattice step, shipped as variants/neural_ca.safetensors. eval_all skips it; README adds the section and updates counts (8 standalone machines, 27-file family round-trip).

neural_reversible: analog read-noise and conductance-mismatch sweeps on the reversible matrix stack. The permutation stays bit-exact through read noise sigma ~ 0.10 (errors at 0.15, where the 0.5-margin error model predicts) and conductance mismatch sigma_G ~ 0.10, the same tolerances neural_matrix8 measures; README states the confirmed tolerance rather than implying it.

neural_reversible: compile the reversible arithmetic to a ternary matrix stack (tools/reversible_matrix.py) and substantiate the no-erasure claim concretely. The 4-bit adder becomes 39 ternary matrices with a Heaviside step; the composed transition is verified a permutation of the state space over all 512 inputs, matches the gate circuit, and clears the analog threshold by the same 0.5 margin as neural_matrix8, so the crossbar realization is bit-exact and information-theoretically lossless. README updated from the by-analogy phrasing to the measured result.

Give neural_attractor and neural_reversible a machine metadata field and eval_all skip entries, so python src/eval_all.py variants/ skips them cleanly instead of erroring (it scores fitness variants and skips standalone machines by that field). README sync: list both new machines in the variant table, intro, and repository layout; correct the standalone-machine count (5->7), the eval_all skip count (four->seven), and the universal-constructor family round-trip (23->26 files, 551->971 MB, both new files codec-verified byte-identical).

neural_reversible: ship the reversible arithmetic core as variants/neural_reversible.safetensors (8-bit in-place adder as a reversible gate sequence plus the Heaviside weights realizing each gate); builder round-trips it and confirms the loaded circuit is a bijection computing b<-a+b

neural_attractor: an energy-based threshold computer where computation is relaxation to a ground state and the program is the coupling matrix. No program counter, no clock, no forward-only execution: clamp any subset of wires and relax. AND/OR/NOT energy gadgets (each zero iff the gate relation holds) make it universal by construction; forward evaluation is exact, and clamping outputs runs circuits backward (an 8x8 multiplier compiled to couplings factors 35=5x7, 143=11x13) or solves SAT. Module, tests, artifact builder, and the shipped coupling matrix.

float mul: full IEEE subnormal support via a unified right-shift renormalizer (right-shift the product by F+1-min(nlz,exp_base)); oracle validated bit-exact against numpy over all classes including deep underflow; all float variants rebuilt

Rename to threshold-computers (heading, citation, URL)

Move source library into src/; repoint module, tool, and README paths

Add neural_matrix8 and neural_subleq8io (universal constructor); move dev scripts to tools/