[SkyeAv]

Second of three stacked PRs for ARS querying. Adds Autonomous Relay System querying to the library, built on @gloriachin's feat/ars-query-support branch, with merge-aware polling ported from @hodgesf's feat/ars-client. Follows the finder conventions landed in #44/#48.

Stacked on #64. The MCP/CLI tool PR stacks on this one.

What changed

• TCT/Query_ARS.py: ARS_neighborhood_finder(node, neighbor_categories, json_file=...) and the new ARS_pathfinder(start, end, intermediate_categories, json_file=...) build one-hop and paths TRAPI queries, submit them to the ARS, wait for every ARA and the merge agent, and return an ARSResult. json_file passes an uploaded TRAPI query through unchanged (edges or paths query graphs). Guangrong's public names are kept: Query_ARS, check_ars_results, fetch_ars_results, format_query_json_forARS_neighborhood.
• ARSResult extends FinderResult with pk, merged_pk, status (an ARSStatus), and summarize(top_n), which returns ranked rows: essence, bound nodes with names, predicates, primary sources, contributing ARAs. isinstance(r, FinderResult) holds, so existing consumers of .raw and .knowledge_graph work unchanged. The merged ARS TRAPI is kept as the ARS returns it, with its own ranking and ARA attribution.
• Inputs resolve like the other finders: names and CURIEs through Name Resolver and Node Normalizer, categories normalized ("Drug" becomes "biolink:Drug"). json_file skips resolution.
• Polling waits for the ars-ars-agent merge child (ARS delay in saving "Done" status on merged result Relay#621) and parses the Python-repr merged_versions_list the ARS actually sends. It uses a deadline instead of retry-count arithmetic and one shared requests.Session.
• config.py: new ars endpoint for prod, ci, and test, selected by TCT_ENVIRONMENT or TCT.configure() (default ci, as before). This replaces the hardcoded URLs from the original branch. The Query_ARS function is not re-exported from TCT, because that would shadow the TCT.Query_ARS module.
• Guangrong's unrelated fixes, as separate commits under his authorship: nodenorm CI-URL removal, informative node-normalizer HTTP errors, CATRAX test-URL fallbacks. A follow-up commit restores DEFAULT_ENVIRONMENT, which the nodenorm change dropped because it was written against an older base.
• notebooks/Test_ARS.ipynb gains pathfinder cells and was re-executed end to end. New docs/source/ars.rst; README tool table updated.

Stats (for audit)

• 9 commits, 14 files, +1900/-10.
• Offline suite: 157 passed, 4 skipped (live tests gated behind TCT_LIVE_ARS=1).
• New offline tests: 20 in tests/test_query_ars.py (Guangrong's uploaded-JSON and notebook-parity cases, plus merge-wait, finder-parity, and summarizer coverage) and 14 node-normalizer routing cases from his branch.
• Live tests: 3 (neighborhood, pathfinder, fixture drift).
• Fixture-drift check: offline fixtures verified as a structural subset of a live trace and merged envelope. During development it caught two real shape facts: the ARS sends merged_versions_list as a Python-repr string, and aragorn/bte merges omit essence and rank.

Performance (old branch vs this PR, same live CI query: asthma, ChemicalEntity)

Measure	feat/ars-query-support	This PR
15 sequential trace polls	4685 ms total, 312 ms mean (new connection each)	1814 ms total, 121 ms mean (shared session), 2.6x faster
End-to-end neighborhood finder	66.9 s, 476 results, 63.9 MB	66.5 s, 476 results, 63.9 MB (parity: time is spent in the ARAs)
Top 20 ranked rows from the merge	no summarizer	0.7 ms on the 476-result merge
Polling limits	15 checks x 30 s, gives up around 425 s	900 s deadline, 15 s interval, waits for the merge agent

Verification

• uv run pytest: 157 passed, 4 skipped.
• Live run 1: TCT_LIVE_ARS=1 uv run pytest tests/test_ars_live.py -v: 3 passed in 199.00 s.
• Live run 2, same command: 3 passed in 227.80 s.
• Manual smoke: ARS_neighborhood_finder("asthma", ["ChemicalEntity"]) returned 476 results and ARS_pathfinder("asthma", "albuterol", intermediate_categories=["Gene"]) returned 1 path. Both reached Done, and isinstance(r, FinderResult) was True.