Shard test files across machines and merge results

Summary

Add deterministic cross-machine test-file sharding plus a merge step for shard results and coverage.

Why

--jobs parallelizes files inside one process, but large CI suites also need to divide work across independent machines without changing which tests run or losing one combined result.

Current behavior

The runner expands an ordered file list and can execute files with local workers. It has no shard identity/count, portable shard manifest, or command that combines independent test and coverage outputs.

Expected behavior

• Accept a validated one-based shard index and total shard count.
• Partition the resolved test-file set deterministically, independently of host count and local --jobs.
• Record enough manifest metadata to detect incompatible or missing shard outputs.
• Merge normalized test results and line/branch coverage without double-counting.
• Derive the combined exit status from all shard results and report incomplete merges clearly.
• Handle empty shards and shard counts larger than the file count predictably.

Scope notes

• Local --jobs remains available within each shard and is not replaced.
• This issue owns deterministic partitioning and artifact merge semantics, not a cloud scheduler or distributed worker coordinator.
• Worker-process pools, intra-file concurrency, and retry/repeat policies are separate concerns.
• The CI-reporting and coverage-threshold result models should be reusable inputs rather than parallel formats.

Start by tracing the runner's ordered file-list expansion and local --jobs execution, then inspect the reusable CI-reporting and coverage-threshold result models. Define deterministic shard manifests, normalized result and coverage merging, and clear incomplete-merge status handling; verify behavior for empty shards and shard counts larger than the file count.