[RBendias]

Summary

Add EnsembleTable.replace_blocks, assuming the ensemble execute API is available. Preserve member schemas and omitted blocks while accepting shared tensor replacements or member-aligned ensemble replacements.

Description of the change

This is an intentionally dependency-blocked extraction from #797: main does not yet provide EnsembleData.execute. The reconstruction path also uses the generic value container introduced by #1367. No execution implementation, runtime fallback executor, type-checker suppression, or existing PR changes are included here.

When all replacement ensembles use the table's member locations, the method delegates to execute(TableTensor.replace_blocks, ...). For differing layouts, it sends only column metadata, omitted blocks, and replacement values to execute(TableTensor, ...); old values of replaced blocks never participate in alignment. All actual alignment remains owned by execute. Semantic tensor types and underlying TableTensor validation are retained.

# Requires the execute implementation.
output = ensemble.replace_blocks(numerical=member_values)

Validation

The candidate method was injected in memory into the real execution backend from #797, without modifying that branch. All 41 selected tests passed: the two new contracts plus Standardize, recipe execution and existing execution regressions. Contract tests cover shared tensors, differently ordered/grouped inputs, simultaneous numerical/datetime replacement, different member schemas and row counts, and preservation of omitted categorical vocabularies.

CUDA measurements used eight members, 50,000 rows, and 32 numerical columns with torch.utils.benchmark.Timer.blocked_autorange(min_run_time=0.2) and peak allocated-memory statistics. Compared with the existing #797 method:

Layout	Existing #797	Candidate	Candidate extra CUDA memory
Shared table, member-specific replacement	652 µs	650 µs	0 bytes
Matching physical batch	523 µs	113 µs	0 bytes
Reordered source batch	592 µs	535 µs	0 bytes

A plain wrapper around execute(TableTensor.replace_blocks, ...) was also tested. It allocated 51,204,608 temporary CUDA bytes in the reordered case because it aligned source values that were immediately replaced. This is why the small location-based fast path and reconstruction path are both present. These are operation timings, not end-to-end estimator speedups.

Not merge-ready on main: the repository type check reports the two missing EnsembleData.execute accesses. All other pre-commit checks pass. On main, the two new integration tests are explicitly skipped until that dependency exists; they are executed against the real #797 backend during manual validation. The dependency failure is not suppressed.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.