docs(sim): rigorous derivation of the evaluation metric

Add METRIC_DERIVATION.md — a standalone, skeptic-facing account of WHY
the metric is regret(linear, sign-split) + effort(direction-split) + an
explicit detection axis, per scenario class, with 3:1 directional weights.

Where THEORY.md is the chronological lab notebook, this is the cleaned-up
proof: each claim is labeled PROVEN (algebra/probability from the plant
identity), EMPIRICAL (named simulation), or VALUES (declared judgment),
and the argument is carried as much by the three falsified hypotheses
(δ²-cancellation, single-scalar E=J_opt/J, sufficiency of ∫e²) as by the
surviving ones. Includes the detection non-recoverability theorem (§7),
the quadratic-blindness lemma (§5.3), the weight-robustness result, the
confirm-debias "−7% gap is optimal" demonstration, and an explicit
falsification checklist (§11) plus an epistemic-status table (§12).

All citations verified against the tree (poisson_floor metrics.rs:864;
commits 31a9dbc / a1d3fa7 / 70fcb26).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: Eric Price