fix(evals): a pairwise judge that errored is not a tie

[asdf8675309]

PairwiseComparison passes when its judge is completely down.

compare() catches every judge error and returns winner: 'tie'. With position_swap on (the default), a judge that never answers produces two ties, which the aggregator scores as (0 + 2*0.5)/2 = 0.5 — and passed = score >= 0.5 makes that a PASS.

The failure is quiet in the worst way: the run reports every pairwise task as passing, and the per-position reasoning strings say "Comparison error: ..." while the score says the output beat the reference. LLMRubric and NaturalLanguageAssert both already return 0 when their judge throws; pairwise is the one model-based grader that does not.

A comparison result gains an errored flag, so a judge that never answered stops being indistinguishable from one that answered TIE, and the scoring moves into an exported pure aggregatePairwise() that returns 0 when any comparison errored. One errored comparison fails the whole grade rather than scoring on the survivor: the swap exists to cancel position bias, so half of it is not a debiased result, it is a biased one carrying a debiased result's score.

Why it's safe: genuine ties are untouched — a judge that answers TIE twice still scores 0.5. The only movement is outage -> 0, so nothing can newly pass. errored is optional, so any existing ComparisonResult literal is unaffected.

Verified by mutation, not just by a green suite. Deleting each load-bearing line in turn, against the 11 tests:

aggregator short-circuit removed -> 7 failures
errored: true in the catch -> 4 failures
errored on the swap arm -> 1 failure

That last one matters: an earlier draft tested only the pure function, and both of the wiring mutations left it fully green — the extracted predicate was covered while the wiring that feeds it was not. Four of the eleven tests drive the real grader through a mocked judge for exactly that reason, and one of them fails the FIRST comparison only, since an all-calls-fail case short-circuits before the swap arm is ever read.

Adds the skill's first test file (11 tests) and a test script. Bare bun test rather than an enumerated directory list, so a colocated test added later is discovered rather than silently skipped.