[Aminsed]

From external review. The overflow case was flagged as unconfirmed — it reproduces, and it is the most serious item here.

Confirmed bug: silent "no outliers"
Coordinates around 1e20 contain no NaN or infinity, so they passed validation, then overflowed when squared. Every selection key tied at infinity and every score came back NaN. fit_predict labelled every point an inlier — including five planted extreme outliers — because NaN < offset is False. An outlier detector reporting nothing wrong when it has failed is the worst available failure mode. Now guarded by a magnitude precondition (checked in logs so the check cannot itself overflow) and a non-finite backstop on the output.

Not a drop-in replacement
score_samples(X)/predict(X) recomputed transductive LOF over X alone — making the new points each other reference population — then applied the offset learned during fit. That is not scikit-learn novelty semantics and it returned a plausible-looking array. The argument is removed, so it is now a TypeError. The claim is gone from the README, the package description and the class docstring, replaced with what is actually covered and an explicit list of what is not.

k complexity corrected
Only the selection scan is k-independent. Neighbour storage is O(n·k) and both the density and score kernels loop over all k, so docs/usage.md saying a large k is "free" and culof.h saying cost does not grow with k were both wrong.

Stale exponents
The README still carried n^2.01 / n^1.13 while docs/implementation.md had the re-baselined n^1.94 / n^1.31 over n ≥ 50,000. My omission when re-baselining.

Benchmark now enforces agreement
It computed relative error and rank agreement, printed them, and asserted nothing. It now raises, so "a speedup on a wrong answer is not a speedup" is executable.

Development Status: Beta → Alpha, since nothing in CI executes a kernel.

50 tests pass (45 + 5 new regressions).