mbr_campaign.STAGES names an aggregate class and a fixed program per stage; both are wrong

[gaberber]

Owner: me (guan). Not a task for anyone else -- filed so the reason is written
down somewhere other than my head, and so the fixes it makes unnecessary
(#2) can point at it.

The table

experiments/qsim/mbr_campaign.py:318-331:

STAGES = {
    "calibration":   (MBRPhaseCorrectionExperiment, EntireFloquetCyclePhaseCalibrationProgram, _calibration_batch),
    "spectrum":      (MBRSpectrumExperiment,        NPhotonHamiltonianSpectroscopyProgram,     _spectrum_batch),
    "propagator":    (MBRPropagatorExperiment,      EncodingPropagatorProgram,                 _propagator_batch),
    "orthogonality": (MBROrthogonalityExperiment,   EncodingOrthogonalityProgram,              _orthogonality_batch),
}

run_stage hands the first column straight to the acquisition runner
(mbr_campaign.py:357):

runner = BatchRunner(station=station, ExptClass=owner, ExptProgram=program, ...)

Both of the first two columns are wrong, and each has already produced a bug:

• Column 1, the class. These are aggregate-analysis classes. Their
  analyze/display need batch_expts, which only the collected object has.
  Submitting them as ExptClass means the worker builds one per child job and
  calls the aggregate methods on a single 1D sweep. Patched at the symptom in
  0f9eddf (worker/runner flag) and a94880b (hasattr guards in display).
• Column 2, the program. The program is not fixed per stage; for
  "spectrum" it depends on whether initial != final. Patched at the symptom
  in d47ab2f (batch.get("program", program)).

Two bugs in one week, from opposite directions, both from this one table. The
table claims a fixed (class, program) per stage, and neither is fixed.

Now vs. target

	now	target
class submitted as ExptClass	the stage class	always the job-sized class, EncodingHamiltonianSpectroscopyExperiment
program	hardcoded per stage in STAGES	taken from the batch object
what a stage class does	acquires and aggregate-analyzes	aggregate-analyzes only; no acquire, no program, never an ExptClass
how the two connect	no seam; role picked at runtime by hasattr or a flag	StageClass.from_batch(runner.execute(...))

Both target columns are already implemented elsewhere in the repo, which is
what makes this a gap and not an open design question:

• experiments/qsim/notebook_helpers/mbr_campaign.py:299 and :504 already
  submit with ExptClass=campaign.EncSpec.
• experiments/qsim/mbr_spectrum.py:1078 already shows
  ExptProgram=spectroscopy_batch.program.
• from_batch (floquet_dark_mode_readout.py:75) exists only to turn an
  acquisition aggregate into a stage class. Its own docstring describes this
  seam. run_stage goes around it.

The archived surface map already recorded both decisions on 2026-09-14:
program_class is the stage key, and new MBR acquisitions record
EncodingHamiltonianSpectroscopyExperiment rather than a stage class. That
doc is marked superseded, so in practice the decision was invisible to anyone
working in these files. Part of this issue is moving the rule somewhere live.

Sketch

STAGES drops both columns -> stage name maps to (analysis class, batch
builder). run_stage becomes:

stage_class, batch = build_stage(...)
runner = BatchRunner(station, ExptClass=EncSpec, ExptProgram=batch.program, ...)
return stage_class.from_batch(runner.execute(batch.configs, ...), station=station)

This removes the need for all three symptom patches above.

Loose ends

• run_stage's off-prod branch (mbr_campaign.py:362-364) also instantiates
  owner(...) per config; it becomes EncSpec too.
• run_stage's return type changes from a raw aggregate to a stage class.
  Callers that already do .analyze() on the result keep working.
• _from_expts uses cls.__new__(cls) (floquet_dark_mode_readout.py:66) to
  skip the acquisition __init__ it should not be inheriting. That can go once
  aggregates are no longer acquisition classes.
• HDF5 filenames change, because they are built from the class name. Handled
  separately as part of the naming pass; not in scope here.

Blocks the jonginn -> main merge.