fix: make IsAdjacentTo symmetric, and normalization with it - #17

The defect

RangeExtensions.IsAdjacentTo switched on the receiver's shape and handled only IFiniteRange<T>; every other shape fell through to false. Its inner switch did handle unbounded operands, so the relation was asymmetric:

(,0].IsAdjacentTo([1,3])  → false     [1,3].IsAdjacentTo((,0])  → true
[4,).IsAdjacentTo([1,3])  → false     [1,3].IsAdjacentTo([4,))  → true
(,0].IsAdjacentTo([1,))   → false

PostgreSQL's -|- is symmetric and answers true for all five — queried directly against the server to confirm, not inferred. The XML doc asserted the broken behaviour as if it were intended, which is why reading the code confirmed the comment rather than catching the bug.

Why it mattered beyond the predicate

RangeSet.From and RangeSet.Union merge neighbours with current.IsAdjacentTo(next) after sorting by lower bound, so an unbounded-start element is always the receiver and always took the broken direction. Sets were built violating the pairwise-non-adjacent invariant the type documents:

RangeSet.From([(,0], [1,)])          was {(,0],[1,)}   now {(,)}
blocks.Union(blocks.Complement())    was {(,0],[1,)}   now the Infinite set

Two sets that should be equal compared unequal depending on how they were built, and a set covering the whole domain did not equal RangeSet.Infinite.

The fix

Rewritten to switch on the pair (range, other), deciding each unordered shape pair once so both receiver orders route to the same test — symmetry is structural rather than something each arm has to remember. Empty, Infinity, and two ranges open at the same end still answer false.

RangeSet.IsAdjacentTo was never affected: it has its own bound-based implementation that tracks infinity flags explicitly.

Verification

Tests were written first and confirmed red before the fix, then green, then red again on reverting only the source:

• symmetry sweep across all 8×8 shape pairs
• the five pairs PostgreSQL was queried about, both directions
• continuous XOR-inclusiveness with unbounded receivers, plus the negative cases
• four normalization consequences, including Union with Complement
• live PostgreSQL: RangeAdjacency_UnboundedShapes_MatchPostgres compares model against server; it fails naming the disagreeing pair '(,0]' -|- '[1,3]' when the fix is reverted

No pre-existing test broke — the fix only adds true where PostgreSQL says true. The existing unbounded tests kept passing because they only ever called the working direction, with comments explaining the receiver "must be finite".

Compatibility

⚠️ Results change for any range or set with an unbounded element adjacent to its neighbour — in every case from a wrong answer to the one the database gives. No API change; PackageValidation passes against the 6.2.0 baseline. Applies to the NodaTime range types, which share the predicate.

Docs: XML doc corrected, README adjacency section gains a symmetric-and-unbounded example with a "Changed in 6.2.1" callout, changelogs for all four packages.

🤖 Generated with Claude Code