[Bug]: relative imports in __init__.py resolve to wrong or empty targets (missing package context)

[orenlab]

Metadata

Field	Value
CodeClone version	2.1.0a1 (reproduced on release/2.1.0a2, commit 932d4b73)
Affected area	Quality metrics / health

What happened?

Relative imports written inside a package __init__.py are resolved incorrectly or silently dropped, corrupting the dependency graph, coupling/cohesion inputs, and relationship facts. This is independent of the src-layout defect (#53) and reproduces in flat layouts too.

Root cause: _resolve_import_target (codeclone/analysis/_module_walk.py:78-90) has no package context. It strips level segments from the raw module_name string, assuming the last dotted segment is always "the module itself". For pkg/__init__.py, module_name_from_path already collapses the name to pkg (the package itself), so the arithmetic over-strips by one level. There is no is_package concept anywhere in the codebase (rg "is_package" → 0 hits).

Proof by executing the unmodified production function:

in pkg/__init__.py (module_name = "pkg"):
  from . import x          -> ""            (falsy target; edge silently dropped
                                             at _module_walk.py:385,410)
  from .sibling import y   -> "sibling"     WRONG — missing "pkg." prefix; edge
                                             points at nonexistent top-level module
in pkg/mod.py (module_name = "pkg.mod"):
  from . import x          -> "pkg"         correct
  from .sibling import y   -> "pkg.sibling" correct

Downstream: _collect_import_from_node (_module_walk.py:377-419) only records a ModuleDep/import name when the target is truthy, so from . import x in any __init__.py produces no dependency edge, no import_names entry, no relationship binding — invisible to build_dep_graph, fan-in/out, cycles, coupling, and MCP implementation context. from .sibling import y records an edge to a module that does not exist.

What did you expect instead?

Relative imports must resolve from package context (package, is_package), not from arithmetic over the raw module string:

• from . import x in pkg/__init__.py → target pkg (plus proven submodule pkg.x when it exists in the module inventory);
• from .sibling import y in pkg/__init__.py → pkg.sibling;
• over-stripped levels (from ... beyond the top) → explicit unresolved_relative observation, never a silently dropped edge.

Minimal reproduction

# repo/pkg/__init__.py
from . import a            # edge pkg -> pkg.a: MISSING today
from .a import helper      # edge should be pkg -> pkg.a; today points at top-level "a"

# repo/pkg/a.py
def helper(): ...

Observed: dependency graph has no pkg -> pkg.a edge from __init__.py; the from .a import helper edge targets a, which is classified external and dropped by _is_internal_target.

Additional context

• Fix contract: Phase 39 spec specs/phase-39-baseline-container-lanes.md §10.2.1 (normative decision table; mandatory package + is_package + node_kind on every module identity).
• Related: [Bug]: dependency graph empty for src-layout repositories (module name canonicalization mismatch) #53 (src-layout canonicalization) — same missing-identity-owner root cause, different symptom; both are fixed by the Module Identity v2 registry.
• Blast: dependency edges, cycles, depth, overloaded modules, coupling classes, relationship/call facts, MCP blast radius and implementation context. Clone fingerprints unaffected (qualname is metadata only, excluded from hash inputs).