fix(ai): stop a single NaN from silently disabling two subsystems

[ApiliumDevTeam]

Third of the stack. Base is #146. This is a behaviour change, not cleanup — that is why it is not in #145.

Why the lint's advice would have been wrong

Both sites carried x.min(hi).max(lo), which clippy wants as x.clamp(lo, hi). They are not equivalent: min/max discard NaN and return the other operand, clamp propagates it.

SurpriseGate

Following the lint would have turned a NaN into a NaN threshold — and every comparison against NaN is false, so should_update stops firing and the memory quietly stops learning. Worse than either bound.

But the real entry point is upstream. record_surprise feeds Welford's algorithm, and one non-finite sample poisons the running mean permanently: mean += delta / n with a NaN delta makes the mean NaN, and every later update keeps it that way. So:

• reject non-finite samples at the boundary, where the damage starts;
• fall back to the configured threshold if unusable statistics arrive some other way — deserialised state is not covered by the entry guard. Not to an invented constant: set_threshold already provides the non-adaptive baseline.

TransactionClassifier::confidence

partial_cmp().unwrap() panics on a NaN distance, which l2_distance produces from a corrupt centroid — so the process died inside a confidence calculation. total_cmp is total and also sorts NaN last, so a corrupt centroid can no longer pass itself off as the nearest one.

And with every distance unusable, the old code fell past second_dist > 0.0 and returned 1.0 — full confidence at the exact moment there is no information. It now answers 0.5, the same neutral value the function already returns when there are no centroids at all.

Verification

Five tests, each pinning a failure that was silent: a poisoned mean, a NaN threshold, the fallback choice, the panic, and full-confidence-from-nothing. Both #[allow(clippy::manual_clamp)] removed.

cargo clippy --workspace --all-targets -- -D warnings exit 0 · 2247 tests green.

The new gate caught two lints in my own test code while writing this, which is the first evidence it works.