Flaky on CI 3.12: graph engine INT128 overflow fails graph-query tests intermittently

[fazpu]

Two different tests in src/tests/surfaces/test_graph_queries.py have failed on CI (3.12 matrix only, so far) with the same underlying fault, each passing on rerun:

• test_s22_document_citation_chain — run 30175604500 (PR Remove unverified claims from binding documents and the retry built on one #142)
• test_a_transient_engine_fault_retries_on_a_fresh_connection — run 30176642324 (PR Render the required array in canonical order, not declaration order #143)

Failure shape, identical in both:

AssertionError: assert Negative(kind=<NegativeKind.BOUNDARY: 'boundary'>,
  explanation='the graph engine hit a transient internal fault on this traversal
  (Overflow exception: INT128 is out of range: cannot add in place.); one retry
  did not clear it', workaround='retry the query, or narrow the hop bound') is None

Notable: the second failing test is itself the test that proves a transient engine fault retries on a fresh connection — i.e. the engine threw the transient twice in a row, so the single retry the code grants was not enough on that runner.

Not investigated further; observed twice on 2026-07-25/26, both cleared by rerun. Worth deciding whether the graph adapter should retry more than once for this specific fault class, or whether the underlying overflow is fixable upstream.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GKENhTLJg1HqhbdwCmmkbc

[fazpu]

Flake sighting on 3.14 (previously catalogued 3.12-only): PR #166 run 30345923632, test_a_transient_engine_fault_retries_on_a_fresh_connection — the INT128 overflow fired on the retry too, so the fresh-connection retry did not clear it ("one retry did not clear it" boundary surfaced instead). Rerun expected to clear; noting the version spread and that a single retry is not always enough.