fix(artifact): bound image saves and recover failed publication

[morluto]

Closes #88

Apply one deadline to nonblocking reads from both Docker-save pipes, stream gzip output, and retain a bounded 16 KiB stderr tail. On failure, terminate the owned process group, close the parent pipe ends, and reap the save process. File and byte uploads now reuse the existing attempt-isolated locator scheme so a failed document insert cannot block the next publication.

Reproduced failure	Before	After
Stalled save with saturated stderr; 0.2 s deadline	Still blocked at 2 s; external watchdog required	Fails in 0.206 s
Retry after document insert fails	Three retries collide with orphan object	Next retry publishes version 1
Temporary archive after failed save	Cleanup unreachable while blocked	Removed
Retained stderr	Unbounded	Last 16 KiB; no stderr staging file
Detached child retains stdout/stderr	Hang or cleanup failure	Returns in 1.501 s at a 1.5 s deadline; archive removed

Measured with local subprocesses and real SQLite/filesystem stores. Tests cover stalled stdout, saturated stderr, surviving descendants, compression errors, valid 16 MiB gzip output, concurrent writes, and both file-upload APIs.

Only newly published object locators change. Existing saved locators still load. Failed publication can leave an isolated object; this change makes retry safe rather than adding garbage collection.

Validation: 6,332 unit/protocol tests passed, 13 existing capability skips; plugin API check passed. Independent correctness and maintainability reviews are clean. Docker/cloud integration remains for CI.

Confidence Score: 5/5

The PR appears safe to merge; the earlier publication hang and cleanup findings are addressed.

What we checked:

• Docker save runs without a shell: The save passes the image name as one argument to Docker. The timeout lines only raise an exception.

Summary

Docker image saves now use one deadline and clean up failed archives. File uploads use a fresh object location for each attempt, so a failed document write does not block a retry.

• Docker image saves finish or clean up under one deadline.
• File uploads get a fresh object location for every try.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Start Docker save] --> B[Read stdout and stderr]
  B --> C[Write gzip archive]
  C --> D{Save finished in time?}
  D -- Yes --> E[Publish archive]
  D -- No --> F[Kill and reap save]
  F --> G[Remove temporary archive]

Loading

_{Reviews (3) · Last reviewed commit: "fix(artifact): enforce deadlines for inh..." · Reviewed by Greptile}