v2.4.0

Fixed

• Resource allocation, Theorem 2 (resource_allocation_validator.py): the alignment-gap lower
  bound H(f,c)^2 / (2 * lambda * alpha_bar) was withdrawn in R7 (Zharnikov 2026k) v2.0.0 because it
  is false in general: a uniform founder has zero gap against every cohort. The validator now uses the
  corrected bound ||w(f) - w(c)||^2 / (2 * lambda * alpha_bar), which holds only under uniform costs
  and condition C4 (||w(c)|| <= ||w(f)||, the founder at least as concentrated as the cohort).
  • compute_alignment_gap_lower_bound() returns None when the theorem does not apply, and the
    validator adds a warning naming the cohort.
  • AllocationReport.alignment_gap_lower_bound is now float | None (default None); new field
    theorem2_applies. New helper satisfies_concentration_condition().
• Cohort boundaries, R3 Theorem 2 (cohort_validator.py): BOUNDARY_FUZZINESS_N8_D010 changes
  from .57, the superseded ambient-dimension variant, to 1 - 0.9^7 = .522 per R3 (Zharnikov 2026f)
  v2.0.0. The warning no longer claims sharp boundaries are "geometrically impossible". It states that
  the bound concerns volume near cell boundaries under the uniform null and that reassignment exposure
  (about two-thirds of observers, R3 Proposition 4) is a simulation result.
• Metamerism detector: when no scalar scores are passed, the report now warns that metameric pairs
  were not tested, so an empty list is no longer read as "none found". The score_tolerance
  docstring states its 0-100 sizing.
• Blind-spot thresholds (2% founder, 5% cohort) are documented as engineering heuristics; R7
  Proposition 3 concerns a founder weight of exactly zero.
• Template 07_resource_allocation.yaml example bound recomputed (.043 to .032). Prompt 07
  attributes weight projection to R7 Failure Mode 1, not Theorem 2.

Added

• Tests for the corrected Theorem 2 (C4 failure, non-uniform costs, 2,000 random Dirichlet pairs),
  the R3 constant, and the metamerism "not tested" warning (161 tests).