Replace Euclidean n/260 hardness scalar with a holographic-screen geo…

[jesusvilela]

…metry

The ACAF critic estimated tunnel hardness with min(1, n/260) -- a flat Euclidean ramp in the variable count that discarded the hyperbolic fabric the rest of the pipeline is built on.

A frame-void instance has already fallen to the boundary at infinity: its Poincare radius r = tanh(gyration) -> 1 (measured -- gyration pins at ~14.16 for every tunnel instance, all n, all alpha). At that boundary the radial coordinate degenerates, so hardness is read holographically off the screen's own intrinsic coordinates in _cosmological_hardness:

• horizon (scale) -- 2^n assignments subtend a comoving hyperbolic
  horizon d ~ n ln2, mapped through the boundary as
  tanh(n/N*) -- geometry's own saturation, no min() clamp
• criticality (caustic) -- a sech caustic peaked on the phase-transition
  ridge alpha_c=4.26, decaying for easy over/under-
  constrained cosmoses
• screen gate -- tanh(gyration) confirms the instance is truly at the
  frame-void boundary before charging full hardness

hardness = screen * horizon * (1/2 + 1/2*caustic), in (0, 1).

The gain is behavioural, not only aesthetic: n/260 is blind to alpha, handing the same swarm to critical and to trivially over/under-constrained instances; the screen reserves the full portfolio for the critical ridge and steps easy tails down, returning a core the old ramp wasted. N*=175 preserves the measured ~220-var heavy-tail knee at ~0.85, so the actor staging is unchanged where calibrated. Every verdict stays certified; this proxy only sizes the mixed strategy, it is never a proof (Charter labels in ACAF_NOTE.md).

Tests: +6 geometric-hardness properties (bounds with no clamp, horizon monotonicity, caustic peak/decay on alpha_c, screen-gate discount, onset calibration). Full suite 656 passed.

Also repair pyproject.toml, which was committed as invalid TOML (a botched merge of two versions: duplicate [build-system]/[project]/[project.scripts] tables, project keys re-opened inside [project.scripts]). It failed to parse and blocked all of pytest. Reconciled into one valid file, keeping the later intentional values (author, MIT license text, Alpha status, URLs, keywords) and the essentials only the first block carried (version 0.1.0, pytest pythonpath).

Claude-Session: https://claude.ai/code/session_01Gkw2ccXyb4DEtqFo29ouQD