feat(resolver): WS0-WS6 context-intelligence upgrade (ambiguity groups, inheritance closure, non-destructive same-dir ranking)

Implements docs/plans/2026-08-18-context-intelligence-upgrade-plan.md's
WS3-WS5 (WS0-WS2 shipped in earlier commits; WS6 evaluated and deferred,
see below). Every workstream verified against the plan's own WS0
precision/false-confidence benchmark before and after, with full
`cargo test --workspace` green throughout.

WS3 -- record ambiguity groups instead of silently dropping them:
`resolve_sites_to_edges` now returns `(Vec<CallEdge>, Vec<AmbiguityGroup>)`.
A call site whose surviving candidate set exceeds MAX_CALLEE_CANDIDATES
no longer vanishes with zero trace -- it's recorded in a new
`ambiguity_groups` table (call_site_id, from_path, candidate_group_key,
candidate_count, reason), persisted by both rebuild_graph and
incremental_graph_update. `callers()` gains `unresolved_group_count` plus
a caveat that fires even when direct_count > 0; `reference_impact()`
gains the sibling `unresolved_many_count`.

WS4 -- inheritance closure as a real resolution mechanism:
`build_resolution_context` now builds a depth-level-aware
inheritance/interface closure from `type_relations` (extends/implements),
gated to `resolved`-confidence rows only -- never `textual`, per the
plan's own hard, non-negotiable rule. `resolve_cross_file_type_relations`
was moved to run before `build_resolution_context` in both graph-rebuild
paths (previously ran after resolve_sites_to_edges, making the closure
exactly one pass stale). resolve_via_inheritance_closure walks levels
nearest-first and only trusts a singleton match at the nearest declaring
level, refusing to pick between two same-depth ancestors that both
declare the same method name (e.g. `interface Mixed extends IA, IB`).

WS5 -- same_dir becomes a non-destructive ranker (C/C++ only):
`call_edges` gains `candidate_rank` (0 = preferred/same-dir, 1+ =
alternate). Found during implementation that the plan's own root-cause
text was wrong for Go/Java: both have real, compiler-enforced package
scoping (confirmed by two existing regression tests breaking under the
naive all-four-languages interpretation), so only C/C++ -- which have no
package/namespace-to-directory correspondence at all -- get the
non-destructive treatment; Go/Java keep the pre-WS5 hard filter.
`callers()`/`callees()` sort `ambiguous` by `candidate_rank`.

WS6 -- evaluated against its own WS0-evidence gate, not met: the
corpus's one real false-confidence case (fixture I / D8) is a different
defect (overlay/static-layer reconciliation) than what WS6 targets
(ambiguous-candidate disambiguation via StackGraph's definition_symbol).
Documented as a reasoned defer in the plan doc; recommends a new WS7 for
D8 instead of building WS6 speculatively.

Net WS0 benchmark result across the whole sequence: call_recall 0.75 ->
0.875 (WS5's fixture E recovering a previously-silently-dropped true
target), false_confidence_rate/false_confident_site_rate unchanged
(0.25 / 0.125) -- a clean recall gain with no precision cost.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Author: Your Name

.gitignore

@@ -63,6 +63,10 @@ __pycache__/
 # resolution benchmark's cloned OSS corpora — external repos, not ours to commit
 benchmarks/resolution/corpus/
 
+# WS0 resolution-precision fixtures are hand-authored and committed, but each
+# fixture's own transient index (rebuilt fresh every run) is not
+benchmarks/resolution_precision/fixtures/*/.calm/
+
 # OS / editor
 .DS_Store
 Thumbs.db

benchmarks/resolution_precision/README.md

@@ -0,0 +1,5 @@
+from lib import bar
+
+
+def use():
+    return bar()

benchmarks/resolution_precision/fixtures/A_import_narrows_fanout/caller.py

@@ -0,0 +1,2 @@
+def bar():
+    return "decoy_01"

benchmarks/resolution_precision/fixtures/A_import_narrows_fanout/decoy_01.py

@@ -0,0 +1,2 @@
+def bar():
+    return "decoy_02"

benchmarks/resolution_precision/fixtures/A_import_narrows_fanout/decoy_02.py

@@ -0,0 +1,2 @@
+def bar():
+    return "decoy_03"

benchmarks/resolution_precision/fixtures/A_import_narrows_fanout/decoy_03.py

@@ -0,0 +1,2 @@
+def bar():
+    return "decoy_04"

benchmarks/resolution_precision/fixtures/A_import_narrows_fanout/decoy_04.py

@@ -0,0 +1,2 @@
+def bar():
+    return "decoy_05"

benchmarks/resolution_precision/fixtures/A_import_narrows_fanout/decoy_05.py

@@ -0,0 +1,2 @@
+def bar():
+    return "decoy_06"

benchmarks/resolution_precision/fixtures/A_import_narrows_fanout/decoy_06.py

@@ -0,0 +1,2 @@
+def bar():
+    return "decoy_07"