perf(lsmr): one true-residual check at a tolerance stop, restart from a refuted one

[schroedk]

Stack 7/9 (feature), based on main (#399 merged). Closes the v0.3.0→main additive solve regression from the #291 exit audit (A + Aᵀ + M⁻¹ apply and serial passes at every tolerance stop, ~1.5 iterations per solve).

• A tolerance stop is held to one fused A apply: ‖b − A x‖ against the recurrence's own ‖r_k‖ (fix(lsmr): underflowed norms, Givens update scale, LSMR's own residual estimate #391), slack 100·tol·‖b‖. A refuted stop restarts from its iterate, at most twice, then FalseConvergence.
• Warm exits still report the normal-equation residual relative to ‖Âᵀb‖ (one Aᵀ + one M⁻¹ apply), as the Python stubs promise for the Adaptive hand-off.
• The α₁ = 0 corroboration stays; the singular-preconditioner blind-spot tests at k > 0 go with the Aᵀ leg, so M⁻¹ must be nonsingular.
• A refused stop reports the returned iterate's normal-equation residual, reseeding once from the staged residual (off the converged path).
• The check sees estimate drift, not drift within range(A); the docs and CHANGELOG say so. test: pin the #286 akm_slopes cell against fake convergence #397's cell stays green.

Hot loop, measured before the split, 8 threads, min of interleaved rounds vs v0.3.0: 2M×3FE/10 iters +0.9% (main +18.1%), 8M×3FE/14 iters +1.5% (main +13.5%). Smoke suite 94/94 converged, correctness 8/8.