Stateful encoders: LevelEncoder and RandomProjection

[MichielStock]

Summary

Introduces the stateful encoder layer (AbstractEncoder{HV}): encoder objects hold hypervector state built once at construction and slot in as the first argument of encode, with decode as the inverse. Two members land here.

LevelEncoder — scalars (breaking)

• Replaces and deletes the level / encodelevel / decodelevel / convertlevel family, no deprecation shims (package is unregistered). The old design rebuilt the level set per call, so separate calls produced incomparable encodings (TODO §1.4/§1.4b); the struct builds the level set once, fixing this by construction.
• Two mechanisms by dispatch: a perturbation ladder for all seven types (bandwidth = flip fraction per step, exposed as a keyword) and fractional power encoding for FHRR (continuous, β bandwidth; explicit level count requests a ladder instead).
• decode is nearest-neighbour by similarity (robust); FHRR additionally offers method = :analytic for clean vectors.

RandomProjection — feature vectors

• x ∈ ℝᵈ ↦ nonlinearity(R·x) with a fixed D × d matrix (:gaussian, :bipolar, :sparse_ternary), per-type nonlinearities, and a parametric scalar-or-vector ternary threshold θ (plus a data-driven target_sparsity constructor and rethreshold sharing R).
• Immutable, no fit; supplied-matrix constructor for reproducible/saved projections.
• decode(rp, hv, references) is nearest-neighbour clean-up only — a random projection is lossy, so no fake analytic inverse is offered; calling without a codebook errors.
• FHRR + :gaussian is exactly a Rahimi–Recht random Fourier feature map (similarity ≈ Gaussian kernel with bandwidth β), locked by a kernel-identity test.

Shared machinery

• New internal phase_encode(z, β) is the single FHRR phase-encoding primitive; both LevelEncoder's fractional power path and RandomProjection's FHRR nonlinearity route through it (equality-locked in tests).

Test hardening

• Regression testsets for two known sharp edges: the ternary constructor's positional collision (supplied matrix vs training data — residual square-matrix ambiguity and silently ignored kwargs recorded as TODO §2.6 for a follow-up rename) and BipolarHV's raw-bits / zero-state polarity traps, pinned so the obvious-but-wrong idioms trip red tests.

Test plan

• Full suite: 825 assertions across types/operations/encoding/inference/representations/extension, all passing
• Doctests machine-generated via doctest(fix = true) and verified in a fresh process (same invocation as the CI doctests job)
• runic --check clean on all touched files

🤖 Generated with Claude Code

[cvigilv]

This looks good, I would merge this to main, merge it against #63, update tutorials and then merge that to main. LMKWYT