Add explicit interaction cycles and experimental CartPole challenges

[btgaskin]

What changed

• add a single InteractionCycle contract between world steps and native neural frames
• make mean, instant, and voting readouts explicit embodiment components
• preserve legacy held-input and mean-readout behaviour by default
• add the four Plank CartPole challenge profiles as registered experimental tasks
• add a fixed-design repeated-start evaluation protocol with retained initial conditions and raw trials
• add task status, tags, protocol metadata, and resolved interaction/readout provenance
• keep the active core benchmark roster exactly Tracking and Pong; Wall remains extended and CartPole remains outside the core aggregate
• document the timing model, evaluation boundary, and CartPole challenge ladder

Why

Temporal spike encoding and categorical voting need a clean place in the standard simulation lifecycle. This change separates sensory encoding, neural-frame timing, output reduction, actuator decoding, and trial-level evaluation instead of embedding a CartPole-specific protocol in the task or motor layer.

The CartPole profiles provide intentionally difficult experimental challenge tasks for probing where a node design stops working. They do not redefine the current core benchmark or imply cross-task score commensurability.

Conformance boundary

The Spike-FF-2 schedule and lower-index voting tie rule are checked against the authors' public TENNLab example. CartPole uses Gym's default explicit-Euler dynamics. The Argyle-4 encoder follows the paper's adjacent-bin, nine-spike description, but its frame schedule remains BrainlessLab-specific until an authors-derived trajectory fixture is available.

Validation

• full Pkg.test()
• focused interaction-cycle, legacy CartPole, Plank CartPole, component-catalog, and API regressions
• Astro site build: 81 pages
• git diff --check

[btgaskin]

Closing as superseded by #8.

Verified against agent/minimal-docs-benchmarks: every source file this PR adds is already present there, and the CartPole surface is complete — all eight tasks are registered (:cartpole, :cartpole_hard, :cartpole_swingup, :cartpole_long, and the four :cartpole_plank_* challenges), alongside PlankCartPole.jl, CartPoleVariants.jl, Interaction.jl, test_interaction_cycle.jl, and test_plank_cartpole.jl.

src/run/Evaluation.jl appears only on this branch because #8 relocated it to src/operations/Evaluation.jl. Merging would re-add a file that was moved deliberately and revert PlankCartPole.jl to an older revision.

The roster statement from this PR — core benchmark is exactly Tracking and Pong, Wall extended, CartPole outside the core aggregate — carries forward and is now the stated release scope.

Only outstanding item: the five site/src/content/docs/core/* pages here have no direct counterpart in #8's restructured site, where handbook/* replaced core/*. Tracked separately as a docs check that the interaction-cycle material survived the rewrite.