fix(openhands): protect native shard discovery before macOS CI retirement - #959

Scope

Recognise python3 scripts/validate_shards.py run as repository-root test*.py discovery in the OpenHands resolver's trusted push gate. This is the prerequisite for #957: replaying its current head showed 258 of 321 tests/*.py files losing protection when the macOS workflow was retired; the fixed gate protects all 321 and gives a new module ci_discovered.

• Base commit: a30c2188e4423f05a7448e5f7a3bcfd858e0f5b8.
• Lane: lane:foundation.
• Owned paths touched: blueprints/runtime-workers/openhands/resolver/push_gate.py, its dated evidence/shard-discovery-protection-20261010.json, tests/test_runtime_worker_openhands_push_gate.py, docs/decisions/2026-10-10-openhands-shard-discovery-protection.md, and manifests/evidence.json.

The repository-tree regression keeps the real shard suite step, removes other workflows and injects a new test module. It fails before the fix on main and on exact #957 source/tree 7df900f7dd6f24b20e3d9e54cfd21aba318d75e8, then passes. The existing protected expectation becomes the stronger ci_discovered rule. A MemoryTree regression retains Git-free derivation and rejects aggregate, runner and run-again as discovery calls.

Correctly identifying tests/test_catalogs.py as a discovered test also removes its eight locations from advisory gate-script import tracing. The historical monitoring receipt stays intact; the test asserts this exact correction, every remaining script count and every recorded shape.

SOTA sources

• native-agent-stack, scripts/validate_shards.py:94-96 at a30c2188e4423f05a7448e5f7a3bcfd858e0f5b8: the shipped wrapper delegates discovery to the native loader with test*.py; lines 186-197 establish the current working directory as root.
• python/cpython v3.12.3, Lib/unittest/loader.py:229-233, 344-374 and 419-444: TestLoader.discover root/pattern semantics and package recursion. The existing native runner and existing gate traversal are reused.

Evidence-class table

Receipt: shard-discovery-protection-20261010.json. The source pin is CPython 3.12.3; execution used the installed Python 3.13.16 client. The existing importable-package discovery boundary still applies.

Local commands run

python3 -m unittest tests.test_runtime_worker_openhands_push_gate.RepositoryWorkflowTests.test_validate_shards_alone_protects_a_new_test_module -v
Before: rc 1, one assertion failure (None, None); after: rc 0.
Exact #957 gate/tree replay: one assertion failure, zero errors before the fix.

python3 -m unittest tests.test_runtime_worker_openhands_push_gate tests.test_runtime_worker_openhands tests.test_runtime_worker_openhands_resolver -v
rc 0; 389 tests, zero failures/errors/skips, 71.812s.

python3 scripts/evidence_manifest.py --check
rc 0; 11,333 files. All four changed/new evidence files registered through scripts.host_receipts.register_file.

python3 scripts/validate.py
rc 0; 70 components, 11,333 hashed files, 4 profiles, 239 receipts.

gitleaks git . --log-opts=a30c2188e4423f05a7448e5f7a3bcfd858e0f5b8..HEAD --redact --no-banner
rc 0; one PR commit scanned, no leaks found; installed Gitleaks 8.30.1.

Decision record

2026-10-10-openhands-shard-discovery-protection.md records the sources, measured comparison, monitoring correction and the wrapper changes that would require revisiting this mapping.

Host evidence

Not applicable: the added receipt records local gate integration and synthetic regression fixtures.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags).
• New/changed workflows declare top-level permissions: {} and grant each job only what it needs.
• No secrets are printed, logged or committed; no new required secret was added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

CC Claude read at this head gates landing. #957 rebases onto this prerequisite after it lands.