[davidackerman]

0.7.0: store opens time out. No placement change and no unit table change.

The bug (cellmap-flow's finding). Retry bounded data reads only. Suppose an http server accepts the connection and never answers. Then wa.open and wa.metadata waited on the metadata fetch forever. Reproduced locally: still waiting after 40 s.

The fix.

• Every metadata document read or written, and every store open, now waits under Retry as data reads do. Each is treated as an operation of no bytes, so by default an attempt waits 5 s times its number, for up to ten attempts.
• tensorstore's own request retries still run inside each attempt.
• The policy is the calling function's retries. wa.metadata, wa.detect and wa.open_tensorstore gain retries=; wa.open, wa.open_pyramid and wa.create already took it.
• With a 0.3 s, two-attempt policy, all three raise TimeoutError within seconds. The error names the URL and the attempts.
• Two things are left alone. Listings, because their size is unknown. The open that creates an array, because a second attempt would find it already made.
• This lets cellmap-flow drop the 5 s requests probe in front of its picker's dtype poll.

Also in this PR: spec open question 5 is resolved, with no funlib_attrs. No live consumer reads funlib attributes from cellmap-analyze or cellmap-flow outputs:

• cellmap-analyze 0.6.0 writes OME only.
• cellmap-flow is the only reader of its own funlib keys.
• dacapo isn't fed these outputs.
• cellmap-data and fibsem-tools read OME only.

Open question 11 now points at an OME-shaped array attribute. The spec also notes that cellmap-data reads an OME translation as voxel 0's corner, which puts it half a voxel off.

186 tests pass locally (1 opt-in stress test skipped), and ruff passes.

https://claude.ai/code/session_0141wgF8mkGbjGfBFKfKhp64