Add the adversary: a real deanonymization engine

Seven heuristics over a target's observable footprint — amount
correlation, exit-minus-fee reconstruction, round numbers, timing
windows, thin anonymity sets, repeated amounts, and cadence. None of
them break a proof; they read the two public legs of a shielded flow
and pair them statistically, which is how deployed pools actually leak.

Findings aggregate through a noisy-or rather than a sum, so six weak
signals cannot add up to certainty and one critical signal is not
diluted by a pile of minor ones. Every finding carries evidence a
reviewer can check and a fix the planner can act on.

The demo contrast holds and is asserted: the naive footprint — round
1,000 in, the same amount out twenty minutes later, weekly — scores 100
and EXPOSED with 16 findings; the Aether-managed footprint scores 0 and
SHIELDED with none. Same adversary, only the behaviour differs.

One correction found while building: the managed scenario originally
had gaps of 4.75, 5.4 and 5.4 days and the cadence heuristic flagged it,
correctly. The demo data was wrong, not the detector, so the data was
de-periodised.

Private in-pool actions are carried with amountKnown false rather than
a fabricated amount — inventing one would make the adversary lie in the
target's favour. Public-event attribution by address is deliberately
unimplemented and returns empty rather than guessing at event decoding.

34 new tests, 158 total.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>