[xe-nvdk]

Ran the 0.2.0 SDK against real Arc builds from main (1943fe87, the 2027.01.1 code), fixed what broke, and checked in the live suites so they can be re-run against any build.

Setups tested

Setup	What it covered
Single node: native binary, auth on, enterprise dev license	Delete, governance, query management, audit, reconciliation, retention, CQ, compaction, backup
Edge sync: hub and spoke as separate processes	Register, rotate, enable/disable, sync, bundle export/import
Cluster: 3 writers and 1 reader	Docker, built from source

SDK bugs found and fixed

Bug	Fix
query_measurement retried Arc's governance 429 (5 requests, ~3 s)	Retries 503 only, like the query POSTs
register_spoke(spoke_id) always failed: Arc refuses the default empty name	name defaults to spoke_id
reconciliation.get_run and spoke status/ledger retried a 503 that won't change (no reconciliation / no edge-sync transport), stalling seconds before raising	Raise at once
A non-dry-run delete could be reported as a failure after it ran. Arc may wait 30 s inside a delete (deleteLocalWait, peer flush), equal to the default client timeout. The caller got "Failed to connect" and lost unreachable_files.	Row and database deletes get a 90 s floor; post/delete accept a per-request timeout
A read timeout said "Failed to connect"	Now says the request was sent and may have been applied
cluster.remove_node returned Arc's 200 {"success": false} (no clustering) as if the node had been removed	Raises ArcError; the id is escaped
The pandas/polars ImportError said pip install arc-client[...]	Names arc-tsdb-client

Each fix has a unit test.

Docs corrected from observed behavior

• Arrow dictionary-encoding thresholds.
• The CQ execute end_time limit, and watermark_status on standalone.
• Spoke ledger discovery.
• What a permissions=[] token can't do: run SQL.
• Data deletes must go to the primary writer.
• assign_compactor's extra 409.

Live suites (opt-in, skipped without env)

• tests/integration/test_live_parity.py: single node plus edge sync. 75 passed, 6 edge-sync skips when hub/spoke aren't running, and 2 xfail(strict=True) for the Arc bugs below.
• tests/integration/test_live_cluster.py: 42 passed, plus 1 opt-in remove_node, which passed when enabled. One flaky fixture is fixed: it removed a file from only one node, and Arc's file puller healed it.
• The run commands are in each module's docstring.

Arc issues filed

• Columnar msgpack type conflict returns 500 instead of 4xx (line protocol and rows return 400 partial write) arc#1424: a columnar type conflict returns 500 instead of 4xx.
• Deleted database still listed by GET /api/v1/databases on local storage (get returns 404) arc#1425: a deleted database is still listed on local storage.
• DELETE with allow_unreachable=true still waits the full deleteLocalWait (30 s) before answering arc#1426 (question): allow_unreachable=true still waits the full 30 s.

Not exercised

• Not set up: tiering (needs S3) and MQTT.
• Schedulers and write diagnostics: still open in feat(schedulers): expose scheduler status and controls #27/feat(diagnostics): expose write protocol health and stats #28.
• Cluster states this setup can't produce: real stream claims, a staged compaction job to abort, assign_compactor success (needs a compactor-role node), and unflushed_nodes (shared storage only).

Checks

1087 unit + integration tests pass without env; ruff and mypy clean.