Validate shared-UI backends, reconnect behavior, and consumers

[uermel]

Parent epic: #53

Objective

Validate the migrated thumbnail path on supported backends and in both consuming applications.

Work

• Run focused legacy-v2 and canonical-v3 thumbnail reads on local, S3-compatible, Secure Shell, and ML Croissant stores.
• Test consecutive Secure Shell reads and a forced connection drop while holding a lazy array.
• Run concurrent thumbnail workers around a forced drop; require completion or one terminal callback and prove shared UI adds no retry.
• Leave private filesystem-replacement and cache-invalidation counts to copick core's own tests.
• Run the public portal smoke, a local portal-shaped OME-Zarr 0.5 fixture, and the exact published core direct-reader fixture.
• Measure cold/warm latency, requests, bytes, and peak memory with at least five repetitions per case; retain raw values plus median and range.
• Install the shared-UI alpha into napari-copick and chimerax-copick 2.0 alpha environments.
• Run optional Server Message Block coverage when infrastructure is available without making it release-gating.

Acceptance criteria

• Local, S3-compatible, Secure Shell, and ML Croissant reads pass.
• Reconnect recovery needs no shared-UI retry and produces no duplicate terminal callbacks.
• The portal and exact core fixtures produce nonempty thumbnails through metadata-defined levels.
• The thumbnail path remains a bounded slice and does not materialize a full volume.
• Raw performance evidence is retained and any repeatable regression is explained before release; no arbitrary percentage is invented.
• napari-copick and chimerax-copick alpha smoke matrices pass.
• Server Message Block remains optional and non-gating.

[uermel]

Implementation update from #56:

• deterministic local, S3-compatible, and ML Croissant thumbnail reads are covered;
• the four-worker concurrency regression now runs in ordinary test CI and requires exactly one successful callback per worker;
• SSH reconnect/drop, live portal, and exact-core-fixture checks remain available as opt-in integration tests;
• the unsolicited ecosystem-validation workflow and its self-invalidating pins have been removed;
• the committed benchmark is retained as diagnostic data only and is not presented as proof of bounded reads.

#56 no longer closes this issue. Keeping it open for the remaining remote legacy-v2 cases, full cold/warm/cache/layout evidence, local portal-shaped coverage, and the downstream napari/ChimeraX alpha smoke matrix before stable cutover.