fix(queries): match :EDGE and filter on kind in typed traversals

[r0h1tb]

Description

Eight read sites matched relationship types that no writer ever creates, so
they could only ever return empty — silently, with no error.

batch_upsert_edges writes every edge as a single untyped :EDGE
relationship and keeps the semantic kind as a property:

MERGE (a)-[r:EDGE {id: e.id}]->(b)
SET r += e

The only other relationship types created anywhere are :RELATES and
:CONTAINS_BLOCK, both of which have matching readers and are fine.

This is option A, as you chose on the issue — align the readers to the
writer. Each site now matches [r:EDGE] and filters with WHERE r.kind IN $...,
the shape #50 already moved the call path to. Variable-length traversals use
all(rel IN rels WHERE rel.kind IN $...), matching find_callers/find_callees.

Related Issue

Fixes #61

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature
• Breaking change
• Documentation update

What was broken

file:line	was	now
api/ast_rag_api.py:470	[:INHERITS|EXTENDS|IMPLEMENTS*1..N]	[rels:EDGE*1..N] + all(rel IN rels …)
api/ast_rag_api.py:486	same	same
api/ast_rag_api.py:520	[:OVERRIDES*1..N]	[rels:EDGE*1..N] + all(rel IN rels …)
api/ast_rag_api.py:916	[r:TYPES]	[r:EDGE] + r.kind IN $type_kinds
api/ast_rag_api.py:921	[r:INHERITS|EXTENDS|IMPLEMENTS]	[r:EDGE] + r.kind IN $inheritance_kinds
api/ast_rag_api.py:995	[r:TYPES]	[r:EDGE] + r.kind IN $type_kinds
api/ast_rag_api.py:1019	[r:INHERITS|EXTENDS|IMPLEMENTS]	[r:EDGE] + r.kind IN $inheritance_kinds
services/search_service.py:499	[:CAPTURES]	[r:EDGE] + r.kind IN $capture_kinds

User-visible effect: find_subclasses, find_superclasses, find_overrides
and the type-usage / inheritance counts in the node-detail and impact paths
returned nothing regardless of the graph.

Two details worth flagging yourself

1. type(r) would have silently changed meaning. The paginated inheritance
query returned type(r) as rel_type, which after this change is always "EDGE".
It returns r.kind instead, so the reported reference type still says
INHERITS/EXTENDS/IMPLEMENTS rather than degrading to EDGE. A naive
rewrite of the eight sites would have broken that field.

2. :CAPTURES is corrected but still returns nothing — for a different
reason, which I want to be explicit about rather than imply it now works. The
parsing layer emits no CAPTURES edges and no Variable nodes at all, so
find_lambdas(with_captured_vars=True) reports captured_vars: [] either way.
The query shape is fixed so it starts working the moment the extractor lands,
and EdgeKind.CAPTURES is declared so the kind is no longer a bare undeclared
string. Happy to drop that hunk if you would rather keep this PR strictly to
the sites where edges actually exist.

Option B

Left undone deliberately, and documented in a comment next to the new kind
groups. Emitting real typed relationships keeps traversals natural and lets
Neo4j use relationship-type indexes for the *1..N inheritance walks, but it
needs a dynamic write (APOC or a per-kind MERGE branch) and is a breaking
graph-schema change requiring a re-index. Worth doing deliberately rather than
as a bug fix.

Checklist

• My code follows the code style of this project
• I have added tests that prove my fix works
• All new and existing tests passed (pytest tests/ -v)
• I have updated the documentation accordingly (comments; no doc pages affected)
• I have run ast-rag evaluate --all — not run, see Testing below
• My changes generate no new warnings

Testing

The regression test is static, for the same reason test_query_parameters_bound
is: proving a traversal returns rows needs a populated graph with real
inheritance and override edges. It collects the relationship types the writers
create and the ones the readers match, and asserts the second is a subset of the
first. It is derived from the source rather than hard-coded, so if a writer is
later changed to emit typed relationships (option B), the matching readers stop
failing it on their own. A second test guards the guard — if the scan finds no
:EDGE writer, the subset check proves nothing and it says so.

On this branch with the source changes reverted and the test kept:

$ git stash                       # revert source, keep the new test
$ pytest tests/test_relationship_types_emitted.py -q

E   AssertionError: These queries match relationship types no writer creates,
E   so they can only ever return empty:
E     :CAPTURES matched at ast_rag/services/search_service.py:499
E     :EXTENDS matched at ast_rag/api/ast_rag_api.py:470, :486, :921, :1019
E     :IMPLEMENTS matched at ast_rag/api/ast_rag_api.py:470, :486, :921, :1019
E     :INHERITS matched at ast_rag/api/ast_rag_api.py:470, :486, :921, :1019
E     :OVERRIDES matched at ast_rag/api/ast_rag_api.py:520
E     :TYPES matched at ast_rag/api/ast_rag_api.py:916, :995
E
E   Relationship types actually written: ['CONTAINS_BLOCK', 'EDGE', 'RELATES'].

1 failed, 1 passed

Full suite, Python 3.12 on macOS:

	passed	failed	skipped	xfailed
baseline (main @ 41e48af)	232	0	1	1
this branch	234	0	1	1

Exactly +2, the two new tests. No pre-existing failures on either side.

test_query_parameters_bound still passes, which matters here — it is what
would catch a rewritten query whose new $inheritance_kinds / $type_kinds /
$capture_kinds parameter was not actually bound at some session.run site.

$ ruff check ast_rag/          # All checks passed!
$ ruff format --check ast_rag/ tests/   # 88 files already formatted

ast-rag evaluate --all was not run. It needs a live Neo4j + Qdrant, and
the committed ast_rag_config.json points at a private LAN (#62), so I have no
way to run it from here. Nothing in this change touches indexing, embeddings or
scoring — it only rewrites read-side Cypher — but flagging it rather than
ticking the box.

Additional Notes

Ping me if you would prefer this split per query family, or if you want the
:CAPTURES hunk dropped.

[lexasub]

@good pr, may merge