Remove the DataType family — python types + `*Meta` typed dicts

[coretl]

Split out of #392, which grew too large to land green in one session. #392 delivered the getter/setter half (per its own "Note on size"); this issue is the remainder: removing the DataType family. Absorbs the naming pass (#396), which was previously folded into #392.

Scope

• Remove the DataType family. Replace with python types + *Meta typed dicts: AttrR(float, precision=3, units="deg", getter=...) and A[AttrR[float], FloatMeta(...)].
• Naming pass (was DRAFT: Naming pass (precision, Limits alignment, Array1D/Table hints) #396): precision; nested Limits with inheritance (Display→Control, Alarm→Warning, assert Warning⊆Alarm; unset ⇒ unbounded); Array1D/Table as the hint AND the runtime structure.

Metadata model (*Meta)

• Per-datatype TypedDicts: FloatMeta, IntMeta, StrMeta, BoolMeta, EnumMeta, Array1DMeta, TableMeta (fields = precision:int, nested Limits + inheritance, units, …).
• Attr* constructors overloaded per datatype so **kwargs: Unpack[FloatMeta] is statically checked: AttrRW(float, precision=3, getter=...) OK; AttrRW(str, precision=3) is a type error. The overload keys off the datatype — passed positionally, or inferred from the getter's return type when omitted (the inference machinery landed with getter/setter IO rework; remove AttributeIORef and AttributeIO #392).
• A superset Meta TypedDict (all fields optional) for generic extras (see Example 4 — SCPI device: annotated attributes + per-attribute filler data #405).
• The Attribute stores the resolved metadata (attr.meta) — no DataType object. Repoint every transport/connector that read attr.datatype.precision/.units/.limits/.choices to attr.meta (enum choices come from the python type). This repointing is the widest part of this issue.
• Constructor runtime-validates meta vs datatype (belt-and-braces; the filler does the same for the Annotated path, ControllerFiller — declarative/procedural split #394).
• *Meta module location is provisional pending Define the top-level public API namespace (flat vs nested) #406.

Out of scope

• Everything getter/setter IO rework; remove AttributeIORef and AttributeIO #392 already landed: getter/setter constructor params, Polled, the Update dataclass, the runtime surface rename (.readback/.setpoint/poll()/set()), AttributeIO/AttributeIORef removal, and the setpoint-callback rework (ADR 0020).
• The @attr decorator — that is @attr decorator sugar over getter/setter constructors #397.

Key files (§9)

src/fastcs/attributes/, src/fastcs/datatypes/, plus every transport and the embedded connector for the attr.datatype.* → attr.meta repoint.

Acceptance

uv run --locked tox -e pre-commit,type-checking,tests green; the demo controllers (src/fastcs/demo/) updated and green.

Note on size

Still large — the repoint sweep touches every transport. If it cannot land green in one session, prefer landing the *Meta types plus the Attribute.meta storage as a coherent green PR with the old DataType accessors kept as a shim, and leave a comment scoping the transport sweep — do not leave a broken tree or a half-migrated Closes.

ADR: 0014, 0017, 0012
Parent: #388
Blocked by: #392

[coretl]

Opened #418 for this.

It does the whole issue rather than the "Note on size" fallback: the DataType family is deleted, attr.dtype is the python type and attr.meta the resolved metadata, the Attr* constructors are overloaded per datatype with Unpack[*Meta], the ADR 0017 naming pass (precision, nested NumericLimits, Array1D/Table) is folded in, and every transport is repointed off attr.datatype.*. No DataType shim is left behind.

Two things flagged for review on the PR: the old min_alarm/max_alarm drove EPICS LOPR/HOPR, which are EPICS' display range rather than its alarm limits, so the nested categories now map display → LOPR/HOPR and control → DRVL/DRVH; and arrays vs tables are told apart at runtime by whether the metadata names columns, since both are held as np.ndarray.

---

Generated by Claude Code