Hold converged steps to the minimum time step size too

[hnil]

Both dt floors — --solver-min-time-step and the restart count — are only consulted when a substep fails. A run whose substeps all converge but whose controller keeps proposing smaller steps consults neither: dt shrinks without bound, the clock stands still, and the run never finishes and never fails.

network/NETWORK_MODEL5_STDW_AUTOCHK does exactly this. Every substep converges with 9 Newton iterations, the network never balances, and dt falls past 1e-57 days with the clock frozen at day 143.107 of 152 — past the point where t + dt == t, so no number of steps can finish the report step.

Applying the existing minimum on the accepted path stops it after 357 steps with:

Time step control proposed a step of 9.872E-13 DAYS, below the minimum of 1.000E-12 DAYS,
while every substep converges.  The run is not advancing past 1.431070E+02 DAYS.

The limit is never reached anywhere in the compareECLFiles suite (128 failures before and after, same set — all pre-existing platform diffs on this machine).

Note: reaching this state at all needs #7263's deck to get past its abort; on master the deck stops earlier for the unrelated reason described there. This guard is independent and worth having regardless.

[akva2]

LGTM. I assume there is a reason why this wasn't rolled into setTimeStep?

[akva2]

jenkins build this please

[hnil]

Checked rather than assumed — instrumented setTimeStep_ to log any call below the floor and ran the whole compareECLFiles suite: zero hits. So rolling it in would not make anything currently throw.

Two reasons I still left it at the call site:

• On the failure path checkTimeStepMinLimit_ already fires first, with a message about the solver failing to converge. Same numeric condition, opposite cause — this one is for the case where every substep converges and the run still stops advancing.
• setTimeStep_ is also reached from the TUNING/NEXTSTEP path, where the size comes from the deck. Nothing in the suite drives that below the floor, but a deck could, and it would then abort rather than be clamped.

Happy to move it if you would rather have one gate.