[bugfix] map keys: compare by op:same-key, never coerce across type families

[joewiz]

[This PR was co-authored with Claude Code. -Joe]

Summary

eXist conflated map keys of different op:same-key families that share a lexical value, on constructor-built maps. Per XQuery 3.1 §17.1, map keys compare via op:same-key — a type-group comparison with no key casting: the numeric family interchanges; the string family (xs:string/xs:anyURI/xs:untypedAtomic) interchanges; every other type matches only itself.

Measured on a clean build of current develop (these are the new mapKeySameKey.xqm cases) against the two reference implementations:

Expression	eXist before	eXist after	Saxon-HE	BaseX	Spec
map:contains(map{"12":"x"}, 12)	true	false	false	false	false
map:get(map{"12":"x"}, 12)	["x"]	()	()	()	()
map:contains(map{"2020-01-01":1}, xs:date("2020-01-01"))	true	false	false	false	false
map:contains(map{true():"x"}, "true")	true	false	false	false	false
map:contains(map{1:1}, 1.5e0) (numeric, lossy cast)	true	false	false	false	false
map:contains(map{5:1}, 5.0e0) (numeric family)	true	true	true	true	true
map:contains(map{"urn:x":1}, xs:anyURI("urn:x")) (string family)	true	true	true	true	true

eXist was the lone outlier on the cross-family cases.

Root cause

A map{…} constructor builds a homogeneously-typed MapType that records a concrete keyType. MapType.get()/contains() ran a private convert() that cast the lookup key to keyType (e.g. integer 12 → string "12") before delegating to the underlying map — so a cross-family lookup that was castable matched a key it should not have. The coercion dates to the original 2012 map implementation.

convert() is now redundant. The underlying map's comparator already implements op:same-key (AbstractMapType.sameKey), and NumericValue/string-family hashCode are already canonical (recent commits c608d55a, 6b63ec83), so cross-type-within-family pairs (xs:integer(5)/xs:double(5.0), xs:string/xs:anyURI) already hash into the same bucket and match correctly via sameKey alone. SingleKeyMapType (built by map:entry) already compares with sameKey directly and was correct; removing convert() makes MapType consistent with it.

Why XQTS misses it

qt3 and qt4 place their cross-family distinctness tests (map/contains.xml, map/get.xml: xs:untypedAtomic("12") vs 12) on map:entry, which yields a MIXED-key map where convert() was a no-op → eXist already passed them. Neither suite has a constructor-built cross-family distinctness test. (Worth a small upstream contribution to qt4tests to close the blind spot.)

What changed

• MapType — removed convert(); get()/contains() delegate straight to the op:same-key comparator. keyType is retained only for the getKeyType() API (documented).
• mapKeySameKey.xqm (new) — XQSuite regression tests using the constructor form: cross-family distinctness (string/date/boolean keys vs same-lexical numeric/other lookups) and the numeric lossy-cast case, plus within-family positives (numeric; string/anyURI/untypedAtomic) and same-type self-matches that must not regress.

Coupling: closes the serialization bare-string "backdoor"

eXist serialization parameters are looked up by their exist: QName. A user could previously also pass them as the prefixed string "exist:expand-xincludes" only because convert() coerced the QName lookup to a string. With this fix, QName and string are distinct keys, so only the spec-aligned exist: QName form is honored — a deliberate, spec-aligning behavior change.

Upgrade note: eXist-specific serialization parameters in an options map must be keyed by their exist: QName (map { xs:QName("exist:indent-spaces"): 2 }); the previously-tolerated bare "exist:…" string key is no longer accepted. (declare option exist:serialize declarations are not affected.)

A full-repo scan found no production/app usage of the bare-string form — only tests. Affected tests updated: serialize.xql, fnSerializeNewline.xqm, and the file module sync-serialize.xqm — prefixed-string cases now assert the key is ignored; the real functionality is retained via the QName form.

Test plan

All on a clean build of this branch off current develop:

• MapTests — 155 run, 0 failures (9 new cases were failing before the fix)
• XQuery3Tests — 1071 run, 0 failures (covers serialize.xql, fnSerializeNewline.xqm)
• file module FileTests — 53 run, 0 failures (covers sync-serialize.xqm)
• ArrayTests (164), MapTest, BifurcanMapTest — 0 failures
• Codacy PMD on MapType.java — clean

XQTS has no relevant tests.

Spec references

• XQuery 3.1 §17.1 — maps; op:same-key definition: https://www.w3.org/TR/xpath-functions-31/#func-same-key