Invert the config dependency so the registry is the single source of truth.
Each component declares its own pydantic config model (a MethodConfig with a
'method' discriminator + its knobs); the CLI config is generated from
registry x schema, and components are constructed straight from a validated
config instance.

Core (osipy/common/config/):
  - MethodConfig base (extra='forbid')
  - method_union(configs): build a discriminated union from {name: config}
  - construct_from_config(registry, cfg): instantiate the selected component
    from its config (validation and construction share one schema -> a knob
    can't be 'collected but ignored')

DSC slice (proves the pattern end-to-end):
  - SSVDConfig/CSVDConfig/OSVDConfig + DECONVOLVER_CONFIGS/_REGISTRY
  - DSCPipelineYAML.deconvolution is the generated discriminated union;
    te/baseline_frames/hematocrit_ratio are real knobs now
  - compute_perfusion_maps accepts an injected fitter; DSCPipeline auto-builds
    it from config (per-method params actually take effect; previously the
    fitter was default-constructed and svd_threshold was dropped)
  - dump_defaults renders the nested config automatically (recurses models)
  - wizard introspects the chosen method's config model to prompt its params

Gate: 825 passed, 0 failed (core incl. GPU); nested YAML round-trips via
load_config; unknown per-method keys rejected.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Apply the registry-driven config pattern to DCE (osipy/dce/config.py): PK model,
T1 method (VFA fit_method linear/nonlinear now selectable), concentration
(spgr/linear now selectable), and population AIF become generated discriminated
unions in DCEPipelineYAML; DCEPipeline auto-constructs from the validated config.
aif_source (pipeline branch) and fitter (in DCEFittingConfig) stay flat.
dump_defaults renders nested; wizard introspects per-method params.

Gate: 834 passed, 0 failed (core incl. GPU); nested DCE YAML round-trips.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Apply the registry-driven config pattern to ASL (osipy/asl/config.py): m0
calibration, difference method, and a new quantification *mode* (single_pld vs
multi_pld) become generated discriminated unions in ASLPipelineYAML. Multi-PLD
(Buxton + ATT) is now a selectable mode producing CBF + ATT maps. The ASL
physiological knobs (t1_tissue, partition_coefficient, t1_blood,
labeling_efficiency, pld, label_duration) and difference_method now actually
flow into quantification (previously collected but ignored). ASLPipeline
auto-constructs m0/difference/quant from the validated config.

Gate: 845 passed, 0 failed (core incl. GPU); partition_coefficient doubling
doubles CBF; multi_pld yields CBF+ATT; nested ASL YAML round-trips.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Apply the registry-driven config pattern to IVIM (osipy/ivim/config.py): the
fitting strategy (segmented/full/bayesian, per-method knobs incl. bayesian-only
prior_scale/noise_std) and the signal model (biexponential/simplified) become
generated discriminated unions in IVIMPipelineYAML. The simplified model is now
selectable (and a latent dS/df Jacobian bug in it, previously dead, is fixed).
initial_guess is now a real IVIMFitParams field threaded into BoundIVIMModel
(previously set on the config but never forwarded). IVIMPipeline auto-constructs
fitter + model from the validated config.

Gate: 874 passed, 0 failed (core incl. GPU); nested IVIM YAML round-trips;
bayesian surfaces prior_scale, segmented rejects it; initial_guess seeds the fit.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Reflect the new nested YAML (model/t1_mapping_method/concentration/population_aif
for DCE; m0/difference/quantification for ASL; fitting/model for IVIM;
deconvolution for DSC). All 8 real-data CLI pipelines pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Convert all YAML pipeline-config examples (tutorials, how-tos, architecture) to
the new nested registry-driven schema; document newly CLI-selectable
capabilities (DCE nonlinear-VFA + linear-concentration; ASL multi-PLD/ATT mode;
IVIM simplified model + per-method bayesian knobs + initial_guess; DSC per-method
deconvolution params). Add docs/explanation/configuration.md describing the
registry x schema config generation. Fix docs/gen_config_docs.py (it imported
removed flat config classes and broke mkdocs build) to render the discriminated
unions. mkdocs build is clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Add a description to SPGRConcentrationConfig.method so the generated YAML
template shows 'spgr | linear' for the concentration block.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Remove unused internals that are NOT config toggles / public capabilities:
- convolution low-level numerical API (conv/uconv/matrix/deconv + registry);
  keep only the production-used convolve_aif + expconv
- shadowed duplicate DSCAcquisitionParams in dsc/concentration (live one is
  osipy.common.types.DSCAcquisitionParams)
- dead custom exceptions MetadataError / osipy ValidationError (never raised;
  the cli/config ValidationError is pydantic's)

Kept (NOT dead): get_gpu_memory_info (used by BatchProcessor in backend/batch.py),
_reset_gpu_cache (test utility), all list_*() registry helpers, legacy SVD
classes, register_aif, types.FittingMethod.

Gate: 795 passed, 0 failed (core incl. GPU); 8/8 real-data CLI pipelines.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

architecture.md no longer lists fft_convolve / the @register_convolution
registry row (removed in the internal cleanup); convolution example shows the
two surviving functions (convolve_aif, expconv).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The CLI template generator now derives the 'A | B | C' choice annotations from
the schema instead of hand-written description strings:
- discriminated-union discriminator lines (method/model/mode/name) list their
  union members' literals (i.e. the registry keys)
- Literal[...] fields list their allowed values
Convert the remaining flat str enum fields (data.format, output.format,
logging.level, DCE fitter/aif_source, ASL labeling_scheme/label_control_order)
to Literal so they validate + autogenerate too; drop the now-redundant
hand-written validators and '| ' option strings from descriptions. New
@register'd method => its choice appears in --dump-defaults with no edits.

Gate: 795 passed, 0 failed (core incl. GPU); 8/8 real-data CLI; mkdocs clean.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

scripts/verify_local_pipelines.sh hard-codes machine-specific data paths and is
a local-only helper; untrack it and add to .gitignore (kept in the working tree).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Drop the 'problem it solves' section and condense; keep the mechanics
(MethodConfig, discriminator, registry x schema generation, validate+build),
a brief add-a-method note, and the per-modality table.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- ruff-format (pre-commit pinned ruff) reformats a test array that the editor
  formatter left expanded
- remove the stale root asl_config.yaml example (flat pre-rollout schema, no
  longer valid; was already deleted in the working tree)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The option/description comment separator used an em-dash (U+2014); on Windows
the dump_defaults -> write_text -> load_config round-trip wrote it as cp1252
0x97 and failed to re-read as UTF-8 (UnicodeDecodeError). Use an ASCII ' - '
separator and assert templates are ASCII in test_dump_defaults_are_valid so it
can't regress.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>