fix: contract sweep — serde feature, no_std, NaN guards, doc accuracy

[NickFlach]

Verify-triage of all 34 open issues against HEAD, then a grouped fix pass. Do not merge — for review.

The headline: the serde feature never compiled

Five issues (#2 #43 #44 #55 #56) all reported "the serde feature can't serialize X". Every one of those types already carried the correct #[cfg_attr(feature = "serde", derive(...))]. The actual cause was one line in Cargo.toml: serde is pulled in with default-features = false (needed for no_std), which drops serde's Vec/String impls. So PhiNode.connections: Vec<usize> and XiSignature.values: Vec<f32> failed to compile — and that killed the entire --features serde build, taking the Vec-free types down with it.

This stayed invisible because tests/serde_feature.rs is gated on #![cfg(feature = "serde")]. Under plain cargo test it reported running 0 tests, and cargo test --features serde didn't compile. The fix is serde = ["dep:serde", "serde/alloc"]; those 10 tests now actually execute.

no_std was broken too — the earlier detect_hives fix reached for f32::rem_euclid, which core doesn't provide. Added to F32Ext.

Numeric guards

Issue	Was	Now
#58	One NaN phase reached every peer via sin(phases[j] - phases[i]) and overwrote healthy oscillators on the first Euler step	Non-finite oscillators quarantined from both sides of the coupling sum, never integrated
#62	Raw Euler accumulation walked phases past 2π (reported 9.23 rad)	Folded into [0, 2π). sin/cos are 2π-periodic, so no dynamics change — pinned by a test
#41	f32::clamp propagates NaN, so k_base * NaN came back as NaN coupling	Holds last good coupling. (The mean_signal() half was already fixed)
#52	exp(-λt) with negative t is a growth term — clock skew made memories stronger than base amplitude	Age floored at 0
#54	Promised [0, 1] but only had a ceiling, so opposed vectors leaked -1.0 into ranking code	Clamped both ends
#53	Single-node graph reported num_partitions: 0, indistinguishable from empty	1 for one node, 0 only for empty

Docs corrected to match code

The README's usage snippet called KuramotoModel::new(vec![...]) and model.sync(dt, k) — neither exists in any version of this API. Replaced with a working example that is now compiled as a test (readme_kuramoto_example_compiles) so it can't rot again. Also removed the bind/bundle/permute primitives and the eigendecomposition-Φ claim (neither is implemented) and the bridge-level chiral mode (#61 — chiral lives in kuramoto).

NATS contract: reactions carry the shared envelope (#28), dream.end matches KANNAKA.dreams as its own description claims (#38), dependency map points at the singular KANNAKA.exemplar.* subject (#39).

Gates

cargo test --all-features    83 + 10 + 6 passed, 0 failed   (was 71 + 0 + 5)
cargo clippy --all-targets --all-features -- -D warnings    clean
cargo fmt --all --check                                     clean
cargo check --no-default-features                           clean  (was failing)
cargo check --features serde                                clean  (was failing)

Closed separately as already-fixed

18 issues were verify-triaged as already fixed at HEAD and closed with evidence: #11 #27 #30 #31 #32 #33 #34 #37 #40 #42 #45 #46 #47 #48 #49 #50 #51 #57.

Left open deliberately

Four need a call on the metric definition rather than a guess from me:

• compute_differentiation_xi() collapses two-vector differentiation to 0 #35 — compute_differentiation_xi returns 0.0 for any two vectors, because both signals are variances over pairwise similarities and there's only one pair (variance of one sample about its own mean is always 0). Real bug, but the fix is a decision about what two-vector differentiation should mean (cosine distance? Xi-signature distance?).
• compute_differentiation_xi() turns mixed-dimension embeddings into xi=0 #60 — mixed-dimension embeddings read as xi=0 ("no differentiation") rather than surfacing an input error, because cosine_similarity returns 0.0 on length mismatch. Needs a call on whether this becomes a Result.
• ConsciousnessMetrics::unified_xi() ignores the xi field #36 — unified_xi() ignores the struct's xi field. The code matches its own documented formula (Φ × r × K × Ψ); the question is whether differentiation belongs in the product or the name is wrong.
• KuramotoModel::sync() reports perfect order for a single oscillator #59 — single-oscillator sync() reports r = 1.0. Mathematically correct (|e^{iθ}| = 1) and there's an explicit test locking it in, but inconsistent with compute_swarm_phi returning 0.0 for n < 2.

🤖 Generated with Claude Code