feat: 6.3.0 — shape predicates, collection expressions, spans, measure, enumeration and the set/range bridge - #18

Length on every range type. The convention follows the domain: a discrete one
counts its values inclusive of both bounds ([2024-01-01, 2024-01-31] measures 31
days, [1,10] measures 10 integers), a continuous one measures the span. Empty
measures zero and unbounded measures null — different answers that stay
distinguishable. The type follows too: long? for the integer ranges, int? days
for DateRange, TimeSpan? for the timestamp ranges, decimal? for DecimalRange,
and Duration?/Period? for the NodaTime ranges, which separate exact elapsed time
from a calendar quantity. Int64Range is the one type whose count can exceed what
reports it — the near-full domain holds more than long.MaxValue values — so it is
computed in decimal and refused rather than wrapping to a plausible negative.

Values() on the discrete range types only. Declaring it per type rather than as
an extension over IRange<T> makes asking a DecimalRange for its values a compile
error instead of a runtime throw, which is the same bargain the five sealed
variants already make. Validation is eager: an iterator would defer the unbounded
refusal to the first MoveNext and surface it at the foreach rather than at the
call that was wrong.

A bridge between the two type families over the discrete domains: ToRangeSet()
collapses runs of consecutive values ({1,2,3,7} to {[1,3],[7,7]}) and
ToInt32Set()/ToDateSet()/... expand back. The shapes describe the same membership
and differ only in density — a thousand consecutive dates are one daterange or a
thousand-element array, and @> against the range is the cheaper question. Both
directions are client-side; PostgreSQL converts between arrays and multiranges
only through unnest and a custom aggregate.

Clamp(value) on every range, and an indexer on the nineteen value set types,
matching what RangeSet already offered.

IRangeFactory.IsDiscrete, a defaulted virtual static, because the property cannot
be derived from NextValueAfter — that returns null both for a continuous domain
and for the last value of a discrete one. RangeFactoryContractTests holds the two
to agreement and checks that a type claiming a step also canonicalizes closed,
with a discovery floor so it cannot pass by matching nothing.

Each guard verified by seeding its defect: dropping an IsDiscrete override, making
Length measure a span instead of a count, letting the bridge merge across gaps, and
deferring the eager validation each turn the relevant tests red.

Co-Authored-By: Claude Opus 5 <[email protected]>

… label

The parse-rejection change was filed under "## [Unreleased]" and then shipped in
6.2.1 — the tag was cut from a main that already contained it. Consumers got a
changed exception-message shape with no 6.2.1 entry describing it, and the
version-is-documented tests passed throughout, because 6.2.1 did have an entry,
just not the whole one. Folded into 6.2.1 where it belongs, in the root and core
changelogs.

Guarded so it cannot recur: ChangelogConventionTests now fails on any heading
that names no version. Every commit on main is a tag away from publishing, and
the release workflow ships whatever Directory.Build.props names, so a heading
meaning "not yet" is false the moment someone tags. Verified by putting an
Unreleased heading back and watching the test name it.

Added "What runs where" to the README: one table for the translated surface with
its PostgreSQL operators, one for the client-side operations with the reason each
cannot translate. The information existed but was scattered across six places in
prose, and the 6.3.0 additions made the client-side half large enough to be worth
stating in one place. The EF package README gains the same client-side table,
since that is where an EF consumer looks.

Fixed a count that contradicted itself: collection expressions were described as
supported by "fifteen value set types" in one place and nineteen in another. It is
nineteen — fifteen closed types plus four wrapper arities. Every remaining count
claim was cross-checked against the code.

Co-Authored-By: Claude Opus 5 <[email protected]>