fix: preserve calls edge direction in graphify query output

[Yyunozor]

Summary

graphify query can print a calls edge backwards — callee --calls--> caller — depending on which endpoint the query term matches. The graph on disk is correct; only the query CLI's text output is wrong.

Reproduction

alpha.py:
    from beta import process_payload
    def run_pipeline(payload):
        return process_payload(payload)   # alpha calls beta

beta.py:
    def process_payload(payload):
        return {"status": "ok", "payload": payload}

$ graphify extract . --code-only
$ graphify query "run_pipeline"      # seeded on the CALLER
EDGE run_pipeline() --calls [EXTRACTED context=call]--> process_payload()   # correct

$ graphify query "process_payload"   # seeded on the CALLEE
EDGE process_payload() --calls [EXTRACTED context=call]--> run_pipeline()   # backwards

graph.json has the correct source/target for this edge in both runs.

Root cause

query loads the graph as an undirected nx.Graph — deliberately, so BFS/DFS can reach both callers and callees of the seed — but _subgraph_to_text renders each edge as EDGE {u} --> {v} in traversal visit order, which on an undirected graph isn't the stored (source, target). Seeding on the callee visits the callee first, so the edge prints backwards.

Forcing directed=True on load (what path and explain do) isn't the right fix here: on a DiGraph, G.neighbors() returns successors only, so a query seeded on a leaf/sink node finds zero neighbors instead of its callers — a recall regression. #829 fixed this bug class for path, explain and the MCP tools, but the query CLI kept the backwards rendering.

Fix

Same _src/_tgt pattern build.py already uses when direction has to survive undirected storage:

• graphify/cli.py: stash each link's on-disk source/target on its edge data as _src/_tgt.
• graphify/serve.py (_subgraph_to_text): render EDGE lines from _src/_tgt when present, falling back to (u, v).

Traversal is unchanged (same node counts per seed before and after); no schema change; path, explain and the MCP tools untouched.

Tests / validation

Two regression tests in tests/test_query_cli.py seed the same calls edge from both endpoints; the callee-seeded one fails on the prior code with the exact symptom above.

uv run pytest tests/ -q
3523 passed, 3 skipped, 0 failed

Parent commit (abff1b1), same environment: 3521 passed, 0 failed — the delta is exactly the 2 tests this PR adds. ruff check clean.

[Yyunozor]

Landed directly on v8 as 4ace951, with the extra guard and tests in 53d0f88 — closing this as superseded. Thanks for the quick review and the hardening!