[DeusData]

What

Kotlin member calls through a class property — a primary-constructor val, a body property, or this.prop — now bind to the property's declared type, on both the sequential and the parallel resolver. Fixes #1555.

Why it was wrong

The cross-file Kotlin walk resolved a call on a typed method parameter by type, but a call on a class property fell to the name guess. The cross registry held each class of the current file from its extraction defs, and their properties carry no type (constructor vals none at all). The walk's typed AST registration of the same class was added behind that entry under the same QN, so every lookup returned the field-less one. Constructor injection is the common Kotlin dependency shape, so repository.inserir(...) on an injected port was bound to whichever same-named method the guess found, including an implementation of another port.

The walk now registers the file's own declarations into an overlay that is chained to the cross registry and checked first. The overlay is indexed once the top-level declarations are collected. The sole-definer fallback walks the whole chain, so its candidate set is unchanged.

Typing many more receivers exposed resolution gaps that untyped receivers had hidden. Each would otherwise have turned a newly typed call into a wrong name guess, so each is handled in this PR:

Shape	Before	Now
member vs same-named extension in the same file	extension could win	member wins (Kotlin resolution order)
x.hashCode() on a receiver typed by another file	per-file kotlin.Any placeholder and cross-pass override both claimed the occurrence → ambiguous → name guess	placeholder dropped when the cross pass resolves the occurrence
count.toString() on Int / String / JDK types	seeded stdlib member with no graph node → name guess	the kotlin.Any builtin (external, no project edge)
property read on a typed receiver whose property QN the graph spells differently (nested class)	fabricated CALL_REFERENCE to a same-named function elsewhere	falls back by name to non-callable targets only → the property's own node
private val handshake: Handshake? next to override fun handshake()	the value shadowed the member → name guess	a value without invoke yields to the member function
bare call to an enclosing-class member inside repo.let { … } or a member extension	resolved against the lambda receiver only → name guess	the enclosing class is checked next (implicit-receiver chain); a call naming a class stays a constructor call

Evidence

Measured on 14 open-source Kotlin repositories (base = current main), indexed with and without the change. The changed edges were hand-judged against the source in a stratified sample: 20 per strategy stratum, 366 edges.

Per edge type (parallel resolver, all 14 repositories):

type	base	this PR	correct gained	wrong added	wrong removed	correct lost
CALLS	169,089	169,349	2,702	5	2,420	27
CALL_REFERENCE	407	295	29	0	141	0
USAGE	292,816	293,119	303	0	0	0
TESTS (derived)	15,128	15,120	54	10	72	0

Another 9,817 CALLS keep their target and are now resolved by the Kotlin resolver instead of a name guess.

CALLS per repository (estimates from the judged strata):

repo	added	removed	correct gained	wrong added	wrong removed	correct lost	dTP	dFP	unjudged
InsertKoinIO/koin	27	33	27	0	31	2	+25	-31	0
JetBrains/Exposed	182	127	182	0	121	6	+176	-121	0
Kotlin/kotlinx.serialization	91	83	91	0	82	1	+90	-82	0
android/nowinandroid	83	81	83	0	81	0	+83	-81	0
apollographql/apollo-kotlin	260	219	260	0	218	1	+259	-218	0
arrow-kt/arrow	32	31	28	4	26	5	+23	-22	0
detekt/detekt	37	36	37	0	36	0	+37	-36	0
google/ksp	70	17	70	0	16	1	+69	-16	0
gotson/komga	138	127	138	0	127	0	+138	-127	0
home-assistant/android	1120	1105	1120	0	1097	8	+1112	-1097	0
lysine-dev/okhttp	448	383	448	0	382	1	+447	-382	0
sqldelight/sqldelight	79	80	79	0	80	0	+79	-80	0
square/moshi	8	10	8	0	9	1	+7	-9	0
square/wire	132	115	131	1	114	1	+130	-113	0
total	2707	2447	2702	5	2420	27	+2675	-2415	0

Judged strata (op × strategy; C = correct, W = wrong, U = undecidable from the source):

op	strategy	edges	C	W	U	P(correct)
+	lsp_kt_method	2144	20	0	0	1.00
+	lsp_kt_any	242	20	0	0	1.00
+	lsp_kt_super	160	20	0	0	1.00
+	lsp_kt_this	84	20	0	0	1.00
+	lsp_kt_operator	66	20	0	0	1.00
+	suffix_match	5	0	5	0	0.00
+	lsp_kt_extension	2	2	0	0	1.00
+	lsp_kt_lambda_it	1	1	0	0	1.00
+	lsp_kt_constructor	1	1	0	0	1.00
+	lsp_kt_top_level	1	1	0	0	1.00
+	lsp_kt_iterator	1	1	0	0	1.00
-	field_type_hint	1528	0	20	0	0.00
-	suffix_match	728	0	18	2	0.00
-	same_module	144	0	15	5	0.00
-	lsp_kt_operator	24	20	0	0	1.00
-	lsp_kt_method	13	0	13	0	0.00
-	lsp_kt_any	6	0	6	0	0.00
-	lsp_kt_this	3	3	0	0	1.00
-	lsp_kt_lambda_it	1	0	1	0	0.00

• Determinism: two runs per repository, identical edge sets on every edge type (14/14).
• Sequential resolver (6 of the repositories): USAGE and CALL_REFERENCE changes are identical to the parallel ones. CALLS move in the same direction with the same strata. The par/seq symmetric difference shrinks from 7,317 to 5,691 edges.
• Known residuals:
  • == on a project type that declares no equals no longer gets an edge to the kotlin.Any.equals builtin (24 edges; an external target, no project edge involved).
  • 5 suffix-match guesses remain where a bare call sits in an implicit-receiver shape the chain does not cover yet (arrow raise() in member extensions, apollo connectToAndroidComponent in a component lambda).

Tests

tests/test_kotlin_typed_calls.c runs one fixture through both resolvers (CBM_INDEX_SINGLE_THREAD=1 and the parallel path). Every shape in the table above has an assertion with a same-named decoy elsewhere. Each assertion checks the target and, where relevant, that a resolver strategy (not a name guess) produced it. On main the test fails for each of the original #1555 shapes. With only the receiver-chain step reverted, exactly its two assertions fail.

Local legs: macOS arm64, kotlin_typed_calls, kotlin_lsp, complexity, usages and the parallel suites; lint-ci (cppcheck 2.20, clang-format, NOLINT, memory-core).