Completes the stateful AbstractEncoder pair begun with LevelEncoder:

- RandomProjection{HV, T}: fixed D×d projection matrix (gaussian /
  bipolar / sparse_ternary), per-type nonlinearities, scalar-or-vector
  ternary threshold θ with a data-driven target_sparsity constructor,
  supplied-matrix constructor, rethreshold sharing R, and
  nearest-neighbour-only decode (lossy encoding: no analytic inverse).
- phase_encode(z, β): single shared FHRR phase-encoding helper; both
  LevelEncoder's fractional power path and RandomProjection's FHRR
  nonlinearity (random Fourier features) route through it.
- Regression tests for two known sharp edges: the ternary constructor's
  positional collision (supplied matrix vs training data; the residual
  square-matrix ambiguity and silently ignored kwargs are recorded as
  TODO §2.6, fix is a follow-up rename) and BipolarHV's raw-bits /
  zero-state polarity traps, pinned as behaviour-locking tests so the
  obvious-but-wrong idioms trip red instead of shipping corruption.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Resolves the src/encoding.jl conflict by dropping the level/encodelevel/
decodelevel/convertlevel family (deleted by the encoders work) while keeping
this branch's docstring reformatting for every surviving combinator.

Fallout fixed:
- api.md listed the four removed functions in an @docs block; restructured into
  Combinators / Encoders sections matching the primitives -> combinators ->
  encoders taxonomy, and added decode, AbstractEncoder, LevelEncoder,
  RandomProjection and rethreshold.
- iris.jl used all four removed functions, and also defined local helpers named
  encode and decode which now collide with the package's exported ones. Ported
  to a single LevelEncoder and renamed the helpers to encodeflower/decodeflower.
  The prose now names the concept it was already explaining by hand: a stateful
  encoder is shared state, which is what makes encode and decode consistent.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Keeps this branch's heading hierarchy and refined prose, and takes two things
from the other draft: the HDC-concept/kitchen-analogy table in the opening,
which orients the reader before any code, and the closing wrap-up, which this
version lacked entirely.

Also updates the 'more encoders' admonition, which still advertised the removed
level encoders, to name the combinator/encoder split and the new encoders.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

New Literate tutorial giving the encoder layer its first documentation: the
token path, the KMer/NGram distinction (with a worked equivalence to the
combinators each is built from), Sequence/BagOfSymbols, and short sections on
LevelEncoder and RandomProjection pointing at the dedicated tutorials.

The headline example is language recognition from character k-mers, ported from
the HDC workshop: seven Wikipedia snippets, one encode call each, and the
similarity ranking recovers real linguistic structure (English/Simple English,
Dutch/West-Flemish, then the Romance and Germanic pairs) with cross-family pairs
at the bottom. This is the payoff for issue #53 and the clearest demonstration
of why KMer is its own strategy.

Also documents the extension point with a working custom strategy.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Gives RandomProjection its first tutorial coverage, built on the workshop's
colour example:

- encoding RGB triples so perceptual closeness becomes hypervector similarity
- blending by bundling
- decode as clean-up against a codebook, with the reason it cannot be inversion
- an associative memory learning each category's average colour via
  bind/bundle/unbind, and the same thing again with two-thirds of every
  observation replaced by noise (multiple-instance learning)
- the tuning knobs: matrix flavours, theta, and rethreshold sharing the matrix
- the kernel connection: FHRR + gaussian is random Fourier features, verified
  numerically against exp(-b^2 d^2/2) in the page itself
- a plain-Julia perceptron sharpening the yellow/green boundary (0.85 -> 0.96),
  explicitly flagged as not-yet-package-API

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

index.md pointed at the wrong repository and its @contents block referenced a
non-existent examples.md, so the manual section rendered empty. Replaced with a
real landing page: what HDC is, installation, a runnable first taste, the
three-layer taxonomy as a table, and links to every tutorial.

Also adds the developer guide to the nav (it was built but unreachable) and
passes repo as Remotes.GitHub, which silences Documenter's 'unable to determine
repository root' navbar warning.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Second act for the iris example: the same nearest-prototype workflow, but with
the four measurements encoded by a single RandomProjection instead of the
key-value + level construction. Scores comparably (0.96 vs 0.93) from an encoder
that needs no domain knowledge, which sets up the real trade-off: the key-value
representation can be decoded feature by feature, a random projection cannot.

Includes a measured warning about standardising. The usual advice is to
standardise before projecting, but on iris that makes things markedly worse
(0.83 +/- 0.06 vs 0.98 +/- 0.03 raw, raw winning 29 of 30 splits) because the
features already share a unit and the largest-spread ones are the most
discriminative. Softened the same blanket advice in CLAUDE.md to match what the
data actually shows.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Literate passes '#hide' through to the markdown, but Documenter still renders
the value of the block's last expression, so Random.seed!(42) was printing a
stray Random.TaskLocalRNG(). Terminated with a semicolon and a hidden nothing.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ts chance level

Two small API additions, both prompted by questions that turned out to have
teeth.

1. Bundling is m-way, not pairwise. For BinaryHV, BipolarHV, RealHV and FHRR the
   bundling rule depends on how many inputs there are, so bundling is not
   associative. Chained 'x + y + z' was already safe (Julia parses a chain of +
   as one variadic call, which reaches bundle with all three), but 'sum(hvs)',
   'reduce(+, hvs)' and '(x + y) + z' fold pairwise and bundle a bundle. Measured
   on five hypervectors, the folded result scores 0.08/0.05/0.14/0.26/0.50
   against its inputs where a true bundle scores ~0.37 for all five: the defining
   property of bundling is lost. 'sum' now bundles the whole collection; the
   remaining pairwise spellings are documented as wrong in the bundle docstring.

2. A similarity score is uninterpretable without knowing the metric: 0.35 means
   'unrelated' under Jaccard (BinaryHV, GradedHV) and 'clearly related' under
   cosine. Adds similaritymetric(HV) and, more usefully, chancesimilarity(HV) —
   the score of an unrelated pair, which is the number you actually need. Both
   are locked by a test asserting the documented baseline matches what random
   hypervectors really score, and the encoding tutorial now teaches reading a
   similarity against its chance level.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- Runic-formats the two new tutorials; every other changed .jl was already
  conformant, so the style CI should be clean.
- Aligns the colours tutorial's advice on standardising with what the iris
  tutorial actually measures (it is not an automatic win).
- Stops tracking docs/src/examples/*.md. These are Literate output regenerated
  by docs/make.jl on every build, and they accounted for a third of this
  branch's diff (1431 of 4383 lines) as machine-generated duplicates of the .jl
  sources. Verified the docs build from a clean state with no .md present.
- Gitignores .claude/ (local dev config).
- Syncs CLAUDE.md and TODO.md: the new similarity traits, the m-way bundling
  rule, the tutorial list, and a stale entry still advertising the removed
  level/encodelevel/decodelevel/convertlevel family.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Mechanical only: keyword-argument spacing, comprehension indentation, and one
trailing space.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>