[林SO (Linxiushen)]

Describe your changes

SearchPoint stores each selection as an (index, value) pair, but is_valid() checks the pair itself for INVALID and nested choices. As a result, invalid conditional choices pass the strategy's discard gate. A two-choice search currently yields [None, valid_config] through sample.passes_configs and counts both as suggested samples. The engine already skips the None configuration; this fixes the earlier strategy boundary.

Unpack the selection before applying the existing recursive validation. Add flat and nested search-point controls, preserve ignored/falsy choices, and exercise conditional rejection through sequential/random samplers in joint/pass-by-pass searches.

Validation on base 1647eb2b589ae8a82b2dbe4429e5fe4335c40092:

• The eight new controls give 7 failures and 1 pass before the fix.
• All 31 tests in test/search pass after the fix, with zero skips. Command: python -m pytest --confcutdir=test/search -o addopts= test/search -q --basetemp=<fresh-local-directory>.
• The root autouse fixture is excluded because it creates unrelated ONNX/model fixtures. These tests exercise the actual search implementation; no model, optimization run, service, or GPU was used. The full repository suite was not run.
• Native lintrunner -a --take RUFF-FORMAT,PYLINT and direct repository-config Ruff pass on all three changed files, as does git diff --check. The native Ruff adapter's Windows canonical paths miss the existing test/** namespace exemption; direct Ruff uses that exemption without source suppressions.

Follow-up on the automated review: the newly effective discard path must also honor the global stop criteria. It now checks the same should_stop condition before continuing, preserving the existing post-step iteration boundary behavior. Six additional regressions cover sequential/random samplers with iteration limits 0/1 and a deterministic elapsed-time limit. All six fail on the unchanged submitted head; the final code passes all 37 test/search tests, zero skips, using the command and CPU-only fixture scope above. Selected native formatting/Pylint, direct repository-config Ruff and git diff --check pass for the follow-up files.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass. (The final 37 selected search tests pass; full suite not run.)
• Update documents if necessary. (No documentation change needed.)
• Lint and apply fixes to your code by running lintrunner -a (selected format/Pylint adapters; Ruff checked directly as described above).
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes: Discard invalid conditional choices before yielding optimization search samples.

(Optional) Issue link

Found while inspecting the search admission path. The reporting-only SearchPoint.to_dict() addition in #2591 covers a separate behavior.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

🟡 Changes recommended

Discarding invalid points bypasses the global stop check, allowing searches to exceed max_iter and max_time; the discard path needs stop enforcement and a bounded-search regression test.

1 open issue

• Check stop criteria when discarding invalid search points · New

What changed in this PR

This PR corrects search-point validation so Olive discards invalid conditional choices before yielding optimization samples or counting them as suggestions.

Changes:

• Unpack stored selections before validating values and nested choices.
• Add controls for invalid, ignored, and falsy values.
• Test conditional rejection across sequential/random samplers and joint/pass-by-pass searches.

File	Description
test/​search/​test_search_strategy.py	Tests invalid-choice rejection and sample counts across search modes.
test/​search/​test_search_point.py	Adds flat and nested validation controls, including accepted values.
olive/​search/​search_point.py	Validates selected values rather than index-value pairs.

_{Review effort: Balanced}

[林SO (Linxiushen)]

Reproduced on the submitted head: both sequential and random samplers consumed all 100 invalid points despite the global iteration/time limits. The discard path now checks the existing global stop criterion before continuing, preserving the current post-step boundary behavior.

Added six regressions for both samplers (iteration limits 0/1 and a deterministic elapsed-time limit). They fail against the unchanged submitted production code, and the fix passes all 37 native test/search tests. Ruff, native formatting and Pylint also pass. Local validation covers CPU search/configuration behavior; the unrelated root model-building fixture was excluded with --confcutdir=test/search.