Architectural change of opencast_io, readers and writers for all drivers added.

[mayankchetan]

Feature or improvement description

Restructures the openfast_io Python package from two monolithic files (FAST_reader.py, 3 652
lines; FAST_writer.py, 2 979 lines) into a layered architecture, with no change to the public
API (see backward compatibility below). This restructuring was primarily driven by the need for an MCP to use with OpenFAST Agentic Kit (OAK):

• Layer 1 — openfast_io/io/: one stateless ModuleIO class per module (ElastoDyn, BeamDyn,
  AeroDyn, AeroDisk, InflowWind, ServoDyn + StC, HydroDyn, SeaState, SubDyn, MoorDyn, MAP++,
  ExtPtfm, SimpleElastoDyn), each implementing read(file_path, base_dir) -> dict /
  write(data, file_path, base_dir) against the same fst_vt dict slices as before. Each module
  is independently usable and testable without an OpenFAST binary.
• Layer 2 — openfast_io/drivers/: OpenFASTDriver (full coupled deck: orchestration,
  cross-module path resolution, OutList registry threading), FASTFarmDriver, and ten new IO
  classes for the standalone module-driver input files (aerodyn_driver.dvr,
  beamdyn_driver.inp, HydroDyn/SubDyn/SeaState/InflowWind/MoorDyn/AeroDisk/SED/UnsteadyAero) —
  previously unsupported anywhere in the package.
• Layer 3 — facades: FAST_reader.py / FAST_writer.py retain InputReader_OpenFAST /
  InputWriter_OpenFAST with identical attributes, methods, and fst_vt layout, now delegating to
  OpenFASTDriver and emitting DeprecationWarning on instantiation.

New ancillary subsystems, each independently importable:

Module	Purpose
parsing.py	The shared parsing/formatting primitives (float_read, read_array, …) extracted from the reader, plus fmt_field() — a fixed-width formatter that switches to scientific notation instead of overflowing the field
outlist.py	Set-based output-channel registry with capture_outlist/emit_outlist used by every module for lossless OutList read↔write round-trips (plus OutList API and legacy bool-dict converters)
schema.py	Version-indexed machine-readable parameter metadata (get_param_info('ElastoDyn', 'NumBl')), with a CI drift check against the Fortran Registry (tools/check_registry_drift.py)
validation.py	Cross-module deck checks (blade-count mismatch, removed parameters, missing DLL, file-reference existence resolved against the case directory)
formats.py	fst_vt ↔ JSON/YAML serialization

The systematic write→read→compare testing this structure enables surfaced and fixed several
pre-existing silent data-loss bugs (details in
openfast_io/docs/architecture-redesign.md, which also records the design rationale and a
section-by-section audit changelog):

• HydroDyn NBodyMod=1: additional-matrix reads used 6 rows instead of 6*NBody (silent
  truncation for multi-body decks).
• SubDyn: GuyanDamp matrix parse crashed on comma-separated floats.
• BeamDyn: multi-blade decks with distinct blade files wrote every blade to the same path
  (blade 3's properties silently overwrote blades 1–2); blade output directories were assumed to
  pre-exist.
• OutList round-trips: channels were dropped on read and/or entire channel registries dumped on
  write for most modules; now captured and re-emitted losslessly for every module including
  ElastoDyn's optional nodal (BldNd_BladesOut) section and SubDyn's SSOutList.
• Packaging: pyyaml was imported but never declared, so clean installs of the package failed on
  first import of the YAML helpers.

Related issue, if one exists

Impacted areas of the software

openfast_io/ only — no Fortran source, glue code, registry, r-test, or build-system changes.
Python consumers of openfast_io (WEIS, WISDEM, ROSCO tooling) are unaffected through the retained
facades: same class names, same fst_vt dict keys, same execute()/update()/set_outlist()
behavior. The only behavior changes are the bug fixes listed above — callers reading multi-body
NBodyMod=1 HydroDyn decks or comma-format SubDyn matrices previously received truncated data or
crashes.

Additional supporting information

• Design document: openfast_io/docs/architecture-redesign.md (before/after layout, per-module
  changes, bug-fix provenance, critical self-review with rebuttals, audit changelog §10).
• The legacy monolith implementations are deleted rather than retained as _legacy files; git
  history is the reference.
• FAST_vars_out.py (the legacy channel bool-dict) is retained for compatibility and emits a
  DeprecationWarning on import; target removal is the next major version together with the
  facades.
• Known deferred cleanups are recorded in the design doc (§9.5): unify the remaining per-module
  OutList helper duplication, align base_dir type hints with the ModuleIO ABC, context-manage
  the three writers still using manual open()/fsync, exclude tests/ from the wheel, and
  implement FASTFarmDriver.write().

Generative AI usage

Substantial portions of this feature were developed with Anthropic Claude (Claude Code), under human
direction for all design decisions. Specifically, a multi-agent workflow was used: Claude Fable 5
orchestrated the work (design/spec interview with the human author, implementation planning, and
adjudication of review findings); Claude Sonnet & Claude Opus agents performed the code implementation, the
per-task independent code reviews, and the codebase reconnaissance; Claude Opus performed
whole-branch final reviews; Claude Haiku handled small fully-specified tasks (documentation,
single-file test fixes). Every task was gated by an independent review pass (spec compliance + code
quality) before proceeding.

Co-authored-by: Anthropic Claude claude@anthropic.com

Additionally, OpenAI GPT5.5 & Google Gemini Pro 3.1 were used as adversarial reviewers of the architectural and implementation plans.

Co-authored-by: Google Gemini gemini@google.com
Co-authored-by: OpenAI ChatGPT chatgpt@openai.com

Test results, if applicable

Three tiers, all runnable via pytest from openfast_io/ (unit tiers need no binary and no
r-test checkout; prerequisites that are absent skip with a stated reason instead of failing):

Tier	Command / prerequisite	Result
Unit + synthetic fixtures	pytest openfast_io/tests -q (no prerequisites)	100 passed, 0 failed (180 skipped = r-test/binary tiers)
r-test corpus round-trips (read→write→re-read→compare_fst_vt over the glue-code decks + all standalone driver inputs)	r-test submodule checked out	238 passed, 0 failed
Full integration (read→write TMax=2 s→run openfast→read ASCII+binary outputs→verify)	built openfast + make regression_test_controllers for DLL cases	26/26 non-DLL cases pass; DLL-dependent cases additionally verified in CI

• r-test branch merging required — not required (no r-test changes; integration fixtures
  are the unmodified r-test cases)

• please squash PR before merging