Let fast_class enable classification on its own

[krystophny]

Risk tier

• T3: physics, output behavior, coordinate convention

A previously inert flag combination becomes live and now emits classification output.

Correctness contract

Intended behavior change

fast_class = .True. with tcut <= 0 and class_plot = .False. now enters the
classifier, writes class_parts.dat, and takes the early exit at
classification.f90:370. Previously it traced ordinary orbits and wrote no
classification output, so the flag was silently a no-op.

Behavior that must not change

Every configuration that already set tcut > 0 or class_plot = .True..
classification_enabled() returns the same value for all of them, so the
dispatch, the bminmax cache decision, and the class_parts.dat writer are
untouched there. Both golden records (classifier_fast, classifier_combined)
set class_plot = .True. and are unaffected.

Coordinate / unit conventions

None touched.

Numerical invariants

None touched. No integrator, field, or classifier arithmetic is modified; only
the predicate that decides whether the classifier runs at all.

Tests added

• unit: bminmax_cache_cases.tsv gains a fast_class column plus three cases.
  0 0 F T F is the discriminating one — it flips only because fast_class
  now enables classification. Read by both test_bminmax_lifecycle (Fortran)
  and test_bminmax_python (the pysimple._bminmax mirror, updated to match).
• integration: test_fast_class_driver.py, four namelist cases over the
  classification matrix. Oracle is the classification output itself — a
  classifying run owes a class_parts.dat with one row per test particle, a
  non-classifying run owes none.
• system: none.
• golden record: none.

Verified the tests fail without the source change: 3 of 4 fail on reverted
sources, all 4 pass with the fix.

Golden-record impact

• unchanged

Both classifier golden records set class_plot = .True., where the predicate's
value is unchanged.

Failure modes considered

• Orbits that never resolve. With tcut > 0 an unresolved orbit stopped at
  the cut. Standalone fast_class has no such backstop, so it traces to
  trace_time. This trades "fractal runs on hard orbits" for "hard orbits run
  long" — the right trade when the unresolved fraction is small. Callers who
  want the cap keep setting tcut.
• iclass stays 0 for unresolved orbits, which downstream consumers read as
  topology code 0, i.e. prompt loss. Pre-existing and not changed here, but it
  becomes reachable in a new configuration, so it is worth stating: consumers
  must separate genuine losses via times_lost / exit_code.
• Second reader of the case table. test_bminmax_python parses the same TSV
  and had to be updated with the Fortran test; caught by the suite.

Manual validation

make CONFIG=Fast, then make test TEST="bminmax|fast_class" — 4/4 pass.

make test-fast has 4 failures (test_chartmap_startmode1,
test_spectre_sympl_volume, test_spectre_sympl_crossing,
test_spectre_validation). All four fail identically on clean main at
04b31af with these changes stashed, so they are pre-existing and unrelated.

[slopqueue]

Review verdict: Comment

See review above.