A fit can return its own seed with success=True when the curve is far below the solver's absolute tolerances

[marcohost33-maker]

Found while closing #123 (round 20 of PR #121), and deliberately not fixed there: it is the
same shape as #123 by a different route, and the fix for #123 does not cover it.

What happens

fitting/gls.py::fit_gls_ar1 on a curve such as 1e-40 * exp(-1.3 t): SciPy's least_squares
meets its absolute ftol/gtol immediately, because every residual is already far below them.
It therefore returns the starting value and reports success=True. The layer accepts that as a
converged fit.

Why the #123 guard does not catch it

#123 is closed by a degeneracy check on the curve's flatness
(ptp(y) <= eps * max|y|). This curve is not flat in that sense — its dynamic range is entirely
normal. What is abnormal is its scale. A guard written against flatness cannot see it.

Why this matters

It is the same class as the three findings of round 20 and of the round before it: a result is
certified by a criterion that never examined the thing it claims about. Here the criterion is
SciPy's convergence flag, and the thing is whether a fit occurred at all.

Acceptance / discrimination requirement

• A fit on 1e-40 * exp(-1.3 t) must not return success=True with the seed as its estimate.
• Positive control, mandatory: the same curve scaled to order 1 must still fit normally, and a
  genuinely converged fit must not be rejected. A guard that fails both is not an improvement, it is
  the solver switched off.
• The discrimination must be shown at exactly this pair — take the fix back, the new test goes red;
  restore, it goes green.

Related: #123 (closed, other route) · #122.