TASK-054: warn when the configured timestep exceeds the stability limit

[AdamClemens]

Closes Stage 9's third finding (F3): stable_timestep() has existed since
Stage 5 and no live run reached it.

A 64x64 lid-driven cavity at the shipped timestep: 0.008 is 2.05x its own
stability limit and diverges at step 17, with nothing said beforehand:

16x16 dt=0.008 stable=0.01562 ratio=0.51  max|u|=0.721    ran 150 steps
32x32 dt=0.008 stable=0.00781 ratio=1.02  max|u|=0.875    ran 150 steps
48x48 dt=0.008 stable=0.00521 ratio=1.54  max|u|=0.918    ran 150 steps
64x64 dt=0.008 stable=0.00391 ratio=2.05  max|u|=6.4e+04  diverged at step 17

What this does

build_simulation_state now computes the limit for the mesh it is about to
run on and warns when the configured timestep exceeds it, naming both numbers
and their ratio.

It lives in simulation_run.py rather than bootstrap.py so run, record
and resume all get it -- record is the path a long unattended run uses.

Non-fatal by design: stable_timestep's 0.25 safety factor is conservative,
and 32x32 at ratio 1.02 demonstrably runs to completion. A gate would reject
runs that work.

The fifth test exists because mutation testing found the fourth missing

_characteristic_velocity reads only what the configuration prescribes --
simulation.velocity and the velocity.* entries of each boundary face's
field_values. That prefix filter is load-bearing and was untested:
replacing name.startswith("velocity.") with an unconditional True left
all eleven tests passing, while letting a declared scalar's wall value
(temperature: 300.0) be read as a speed of 300 -- shrinking the limit by
two orders of magnitude and firing this warning on configurations that are
perfectly stable.

That is the same scalar/velocity conflation TASK-052 fixed one layer down,
reappearing in new code that reads the same mapping.
test_a_declared_scalars_wall_value_is_not_read_as_a_speed closes it and
fails under that mutation.

Verification

By hand on all three subcommands against a real 64x64 configuration:

configured numerics.timestep 0.008 exceeds this mesh's own stability limit
0.0039062 (2.05x) -- the run may diverge

Silent at the shipped 16x16 (0.51x). make ci green locally: 1219 passed,
20 skipped, all 17 structural checks.

Stage 9 stays open

Not closed at 7-of-7. TASK-055 is drafted to carry Criterion 3's remaining
half -- BoundaryFaceConfig.velocity is still validated for mutual
exclusivity and zero net flux, then read by no engine code -- and Criterion
6's documentation grep. It carries an open design question (wire the field,
or reject it) that wants a maintainer's answer before implementation rather
than the easier reading picked silently.

🤖 Generated with Claude Code