[CI] Cross-platform — Part 2: ARM/Spark workflow#5698
Open
hujc7 wants to merge 59 commits into
Open
Conversation
![isaaclab-review-bot[bot]](https://avatars.githubusercontent.com/u/157846462?s=60&v=4){kind=small}
There was a problem hiding this comment.
Update (df786d2): Reviewed incremental diff since 5cbba72 (1 file changed, 1 commit).
New commits reviewed
| Commit | PR | Description |
|---|---|---|
| feba370 | #5826 | Defer pxr loading in isaaclab.sim, .assets, .scene, .sim/utils |
| 207f319 | #5844 | Log Kit version info on AppLauncher startup |
| c54e512 | — | Scope test tags to arm_ci only; misc review fixes |
| ea8ecd6 | #5891 | Split tasks into core/contrib |
| acb09f2 | — | Merge remote-tracking branch origin/develop |
| 5cbba72 | — | Move cartpole smoke test into test/core/ after tasks split |
| df786d2 | — | Raise cartpole perception smoke-test timeout for cold shader cache |
Latest change (df786d2)
Raises test_train_cartpole_perception timeout from default (presumably matching state test) to 1800 s to accommodate first-run shader compilation on cold GPU cache. Adds a comment explaining the reasoning. Minimal, sensible change — shader compilation on first camera-enabled run can easily exceed 600 s on some hardware.
Full analysis (unchanged from prior update)
1. Deferred pxr imports (#5826) ✅ — All 9 modules now defer from pxr import ... into function bodies. Type hints stay under TYPE_CHECKING. Fixes Kit startup crash when env-cfg parse pulls pxr before AppLauncher.
2. Kit version logging (#5844) ✅ — Simple _log_kit_version_info() added to AppLauncher. Prints Kit version, kernel version, and git hash to sys.__stderr__.
3. Tasks core/contrib restructuring (#5891) ✅ — Massive but mechanical: direct.* / manager_based.* → flat core.* / contrib.*. Gym IDs unchanged. Docs updated. Changelog fragment present.
4. CI/test adjustments ✅ — Test markers narrowed, cartpole smoke test moved, perception timeout raised for cold cache.
Previous concern status
| Issue | Status |
|---|---|
P1 (conftest.py OSError handling) |
✅ Resolved |
| P2 (TAB delimiter) | ✅ Resolved |
No new issues found. LGTM from bot. 🤖
This was referenced May 20, 2026
Foundation for cross-platform CI. Registers four pytest markers (windows, windows_ci, arm, arm_ci), teaches AppLauncher to recognize them in argv so they do not leak into Isaac Sim's argparse, and moves the AssetConverterBase USD scratch directory from a hardcoded /tmp/IsaacLab to tempfile.gettempdir() for cross-platform compatibility. Tags source/isaaclab/test/deps/test_torch.py and test_scipy.py with the new markers so they are selectable by future cross-platform jobs. Workflow files (arm-ci.yaml, windows-ci.yaml) ship in follow-up PRs.
hujc7
force-pushed
the
jichuanh/windows-spark-ci-cartpole
branch
2 times, most recently
from
00966f4 to
ca8fa07
Compare
hujc7
changed the title
hujc7
changed the title
hujc7
changed the title
hujc7
changed the title
hujc7
force-pushed
the
jichuanh/windows-spark-ci-cartpole
branch
4 times, most recently
from
4b8ecb8 to
3e904c3
Compare
Mirrors build.yaml's spirit but stays minimal for the aarch64 path: Tier 1 (gates none — continue-on-error): general-arm, install-arm, kit-launch-arm Tier 2 (meaningful, marker-filtered): kitless-arm, determinism-arm Every job sets continue-on-error: true while the aarch64 runner setup stabilizes. Every pytest invocation passes --timeout=N --timeout-method=signal so a single hung test cannot consume the whole job slot. Inline scripts use set -e to fail on the first nonzero return. Tags three test_rendering_*_kitless.py files plus test_differential_ik.py and test_operational_space.py with the arm_ci marker so the Tier 2 jobs can select them via pytest -m arm_ci.
hujc7
force-pushed
the
jichuanh/windows-spark-ci-cartpole
branch
from
3e904c3 to
98a5b68
Compare
hujc7
force-pushed
the
jichuanh/windows-spark-ci-cartpole
branch
7 times, most recently
from
b7862d0 to
48819bf
Compare
- source/isaaclab/setup.py: gate pytetwild==0.2.3 with `platform_machine != 'aarch64'`. No aarch64 wheel on PyPI; source build fails because the transitive `geogram` dep hardcodes `-m64` in its CMakeLists. The single call site at sim/schemas/schemas.py already lazy-imports it with a clear "install manually" message, so aarch64 users keep everything except automatic volume-deformable tetrahedralization. - docker/Dockerfile.arm-ci: new lightweight Dockerfile that layers cmake / build-essential / git + EULA env vars + a bash entrypoint onto the multi-arch Isaac Sim base image. Replaces the previous inline `docker run + apt-get install + docker commit` chain in arm-ci.yaml. - docker/Dockerfile.base: reverted the libgmp / libmpfr / libeigen3 / libcgal / libboost additions from the arm64-conditional apt block — those were only needed by pytetwild's fTetWild build, which we no longer install on aarch64. - .github/workflows/arm-ci.yaml: build via docker/Dockerfile.arm-ci instead of the inline apt-install-and-commit pattern. Test steps no longer need to re-specify --entrypoint or EULA env vars on every docker run.
hujc7
force-pushed
the
jichuanh/windows-spark-ci-cartpole
branch
from
48819bf to
b3d7f7c
Compare
Forces run_docker_tests=false in build.yaml's changes job so all heavy test jobs skip via their existing if-gate. Saves CI runner time + cost during ARM CI iteration on PR isaac-sim#5698. Must be reverted before final review.
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
Comment on lines
+736
to
+741
| new_test_files = [] | ||
| for test_file in test_files: | ||
| with open(test_file) as f: | ||
| if marker_token in f.read(): | ||
| new_test_files.append(test_file) | ||
| test_files = new_test_files |
Contributor
There was a problem hiding this comment.
The second-pass scan opens test_file without a try/except OSError, so a race-condition delete, permission error, or unreadable symlink on CI would raise an unhandled exception and abort the entire test orchestrator session. The first scan (pre-scan at line ~703) correctly wraps the same open() in except OSError: continue, but this second scan omits that guard. The inconsistency means a transient filesystem issue during the second pass would produce a hard failure rather than a graceful skip.
Suggested change
| new_test_files = [] | |
| for test_file in test_files: | |
| with open(test_file) as f: | |
| if marker_token in f.read(): | |
| new_test_files.append(test_file) | |
| test_files = new_test_files | |
| if ci_marker: | |
| # Match both `@pytest.mark.<marker>` (per-function) and | |
| # `pytestmark = pytest.mark.<marker>` / `pytestmark = [..., pytest.mark.<marker>, ...]` | |
| # (module-level) by looking for the common `pytest.mark.<marker>` substring. | |
| marker_token = f"pytest.mark.{ci_marker}" | |
| new_test_files = [] | |
| for test_file in test_files: | |
| try: | |
| with open(test_file) as f: | |
| if marker_token in f.read(): | |
| new_test_files.append(test_file) | |
| except OSError: | |
| continue | |
| test_files = new_test_files |
Comment on lines
+66
to
+81
| echo "::error::detect-changes: pattern line missing tab separator: '$line'" | ||
| exit 1 | ||
| fi | ||
| patterns+=("$line") | ||
| done <<< "$PATTERNS_INPUT" | ||
|
|
||
| if [ "${#patterns[@]}" -eq 0 ]; then | ||
| echo "::error::detect-changes received no valid pattern lines" | ||
| exit 1 | ||
| fi | ||
|
|
||
| render_table() { | ||
| local files="$1" entry regex desc count sample | ||
| echo "| Pattern | What it covers | Matched files |" | ||
| echo "|---|---|---|" | ||
| for entry in "${patterns[@]}"; do |
Contributor
There was a problem hiding this comment.
The action requires literal TAB characters between the regex and description in each pattern line, and exits with code 1 if any line lacks a tab (lines 72–73). A git attribute like text=auto, an editor that normalizes whitespace, or copy-paste from a browser would silently convert tabs to spaces. When that happens the changes job fails before writing run to GITHUB_OUTPUT, causing all downstream arm-ci jobs to error rather than run — no test results are produced and branch-protection checks would report a failing status. A spaces-delimited format or a structured YAML list would be more portable.
Comment on lines
+1
to
+10
| # Copyright (c) 2022-2026, The Isaac Lab Project Developers (https://github.com/isaac-sim/IsaacLab/blob/main/CONTRIBUTORS.md). | ||
| # All rights reserved. | ||
| # | ||
| # SPDX-License-Identifier: BSD-3-Clause | ||
|
|
||
| # ARM/Spark CI — exercises Isaac Lab on aarch64 Linux self-hosted runners | ||
| # (NVIDIA DGX Spark). Single job that builds Dockerfile.base for linux/arm64 | ||
| # and runs the `arm_ci`-marked pytest set against it. ECR is not wired for | ||
| # arm64, so build output stays local to the runner; the action's local | ||
| # exact-tag and deps-tag checks short-circuit rebuilds when the runner is |
Contributor
There was a problem hiding this comment.
build.yaml run_docker_tests: 'false' override
The PR description checklist includes an unchecked item: "Revert the TEMP build.yaml run_docker_tests: 'false' override before draft → ready." build.yaml is not part of this PR's diff, which makes it unclear whether the override is already in the base branch (from Part 1 #5695) or still present on this branch. If the override is still active, merging would suppress Docker-based tests in the main Linux CI pipeline.
The pre-scan at the top of pytest_sessionstart wraps its `open()` in `except OSError: continue` to tolerate races (a marker-tagged file deleted between os.walk and open(), an unreadable symlink, etc.). The post-scan filter further down was missing the same guard, so a single transient I/O error during the marker-token check would raise an unhandled exception and abort the orchestrator before any test ran. Mirror the pre-scan's try/except so both passes degrade gracefully. Flagged P1 by Greptile auto-review on commit 4bd088f.
Pre-scan: log loud `::warning::` then continue. The pre-scan is best-effort (missed files may still be picked up via _collect_test_files's normal walk path), so silent continue was acceptable but unhelpful — adding a visible warning lets the issue surface without aborting. Post-scan: raise RuntimeError. This is the final filter pass — a silent drop here would let a marker-tagged test exit the run while the orchestrator still reports success. Loud abort on OSError is the right signal; the failure case (permission flap, race-deleted file, broken symlink on a previously-walked path) is exotic enough that hitting it indicates a real CI environment issue worth investigating. Replaces the silent `continue` introduced in a9711ef from Greptile's P1 suggestion — that suggestion mirrored pre-scan's behavior mechanically but missed the semantic difference between the two passes.
Match the existing ISAACSIM_CI_SHORT post-scan's behavior — no OSError handling, let `open()` propagate. Same semantic for both markers: if a test file can't be read during scan, the orchestrator aborts with the raw OSError and CI goes red. No silent drops. Reverts the try/except additions from a9711ef (Greptile P1 suggestion) and d7fb1bd (the differentiated raise/warn version).
Member
|
Why are we creating yet another separate CI for ARM instead of just letting some of the tests from run on ARM? I vote against that. |
build.yaml's `changes` job had its own ~95-line inline detector while arm-ci.yaml consumed the new detect-changes composite. Route both workflows through the composite so the gating logic stays in one place. Extract the deps-hash computation into _lib/compute-deps-hash.sh, sourced by both docker-build (local-store check) and ecr-build-push-pull (registry check). The two caches stay independent; only the hash schema is shared, which removes the silent-drift risk the docker-build docstring previously warned about. Extract the docker-config + nvcr.io login into _lib/setup-docker-config.sh, sourced by both composites. The script is idempotent (early-returns when DOCKER_CONFIG already points at a valid temp dir), so the `skip-docker-config` input on docker-build and the corresponding plumbing in ecr-build-push-pull's delegation are no longer needed. Bump arm-ci.yaml's actions/checkout from v4 to v6 to match the rest of the workflows.
Per @myurasov-nv's review feedback: same OS (Linux), different runner pool shouldn't justify a separate workflow file. Other multi-arch repos (NVIDIA Warp, NumPy, newton, PyTorch) put cross-arch jobs in one workflow with different `runs-on:` labels. The build/test work itself doesn't change. Add an `arm-ci` job to build.yaml that builds and tests on the same [self-hosted, arm64] runner (no ECR for arm64, so the image must stay local to the runner that builds it). Source-side fixes from the original series stay: AppLauncher EXP_PATH fallback, pytetwild aarch64 gate, OVRTX/OVPhysX skip-on-aarch64, conftest.py CI_MARKER orchestration. Gate the amd64 `build` and `build-curobo` jobs on absence of the `ci-skip-amd64` PR label so this branch can iterate on arm-ci alone without burning GPU minutes. All 18 test-* jobs skip transitively via `needs: build`. Remove the label to re-enable the full amd64 surface. Delete .github/workflows/arm-ci.yaml.
…park-ci-cartpole # Conflicts: # .github/workflows/build.yaml # source/isaaclab/setup.py # tools/conftest.py
hujc7
added
ci-arm-only
Iteration helper. The label-gate added in the prior commit (then named ci-skip-amd64) only skipped build.yaml's amd64 build jobs. Expanding to also gate docs.yaml (build-latest-docs, build-multi-docs), install-ci.yml (both x86 and arm installation test jobs), and license-check.yaml so that PRs iterating on a single platform can run only their target workflow. Rename ci-skip-amd64 -> ci-arm-only to reflect the broader scope.
hujc7
added a commit
to hujc7/IsaacLab
that referenced
this pull request
Jun 1, 2026
PR 5695 lays foundation only. The two tagged tests (test_scipy.py, test_torch.py) had pytestmark applied here but no in-PR consumer — the arm-ci and windows-ci workflows that select on these markers ship in isaac-sim#5698 and isaac-sim#5700 respectively. Each downstream PR should land the marker annotations on the tests it actually wants in its canary, alongside the workflow that runs them. This keeps 5695 self-contained as scaffolding and prevents unused annotations from accumulating here.
After folding arm-ci.yaml into build.yaml, the detect-changes composite had a single caller. The composite + sparse-checkout pattern was overhead to support cross-workflow sharing that no longer exists. Restore the inline bash detector that was on develop before the migration. Net: -50 lines, one fewer directory under .github/actions/.
arm64 runs on a different runner pool ([self-hosted, arm64]) than the
amd64 build ([self-hosted, gpu]), so the local docker tag could not
collide in practice. The extra env var was defensive only.
Follow the same suffix pattern that build-curobo already uses
(`${{ env.CI_IMAGE_TAG }}-curobo`) for consistency.
`local-hit`, `local-deps-hit`, and `was-built` were declared as action-level outputs anticipating downstream telemetry callers that never materialized. Grep across .github/ shows zero consumers. Internal step outputs of the same names continue to gate the conditional steps inside the action (the local exact-tag short-circuit, the local deps-tag short-circuit, the post-build deps-tag application), so the deletion is purely surface-area: external declarations go, in-action flow unchanged.
hujc7
removed
the
ci-arm-only
Cleanups in response to self-review: 1. compute-deps-hash.sh refactored from source-with-env-vars to exec-with-positional-args. Caller captures hash via $(...) or pipes into $GITHUB_OUTPUT. More idiomatic, testable from any shell, no env-var contract. 2. docker-build gains a `deps-hash` input. ecr-build-push-pull computes the hash once in its registry-side check step, writes it to $GITHUB_OUTPUT, and passes it forward to docker-build's delegation. docker-build's local check uses the input if non-empty, else computes itself. Hash now computed exactly once per build. 3. `extra-tags` input removed from docker-build. The single consumer (ecr-build-push-pull) now does its own `docker tag <commit> <ECR>` step after the build returns. Less surface area on the primitive. 4. `evict-stale-cache` input added on docker-build (default false). The 14-day eviction is opt-in rather than implicit. No current caller opts in; runners that need disk cleanup do so explicitly. 5. Three unused action-level outputs (local-hit, local-deps-hit, was-built) on docker-build removed — no consumers grep'd. 6. docker-build description reverted to develop's one-liner. 7. ci_marker pre-scan in tools/conftest.py removed. The pre-scan previously overrode TESTS_TO_SKIP for marker-tagged files, which only papered over one mismatch (test_differential_ik.py is in TESTS_TO_SKIP as "Failing" yet was arm_ci-tagged). Respect the global skip list: a skip-listed test stays skipped on all platforms. arm-ci canary goes 8 -> 7 with the one known-failing file correctly dropped. 8. ci-arm-only label gates removed from build.yaml, docs.yaml, install-ci.yml, license-check.yaml. The iteration helper was one-shot; arm-ci is now green, the gate's job is done. Drop the gates and the corresponding repo label. 9. triggered_jobs string in build.yaml's `changes` job reverted to develop's text — the "non-root verify is folded" detail snuck in from an earlier commit and isn't strictly relevant.
The five pre-existing inputs (image-tag, isaacsim-base-image, isaacsim-version, dockerfile-path, context-path) had their descriptions rewritten during the split out of ecr-build-push-pull, with no behavior change. Restore develop's original wording; keep descriptions only on the genuinely-new inputs (platform, cache-from, cache-to, deps-hash, evict-stale-cache).
F2: test_differential_ik.py carried `pytestmark = pytest.mark.arm_ci`,
but the file is in tools/test_settings.py TESTS_TO_SKIP ("Failing"), so
_collect_test_files drops it before the ci_marker filter ever sees it —
the marker was inert. A skip-listed test should not claim to run on any
platform; remove the marker. (The other arm_ci-tagged tests are not in
TESTS_TO_SKIP and are unaffected.)
F3: the changelog fragment credited a pytetwild aarch64 exclusion, but
that gate now lives on develop (via the packaging cleanup that removed
setup.py) and is already documented by a separate fragment. This PR's
diff no longer touches pytetwild gating, so the bullet described a
change not present here. Drop it; keep the in-scope AppLauncher
EXP_PATH fallback entry.
The two _lib helpers were a sourced bash script (setup-docker-config) and an exec'd bash script (compute-deps-hash) — two different calling conventions for what are both "shared step-logic." Unify them on the GHA-native pattern: each becomes a composite action with an inline `run:` block and a declared interface. - _lib/compute-deps-hash/action.yml: inputs (dockerfile-path, isaacsim-base-image, isaacsim-version) -> output `hash`. Both docker-build and ecr-build-push-pull `uses:` it, so the deps-hash schema lives in exactly one place (no drift between the local-store check and the registry check). - _lib/setup-docker-config/action.yml: no inputs; writes DOCKER_CONFIG to $GITHUB_ENV and logs into nvcr.io. Idempotent (a second invocation in the same job short-circuits), so ecr-build-push-pull delegating to docker-build doesn't redo it. Call sites change from `run: . _lib/x.sh` / `$(_lib/x.sh ...)` to `uses: ./.github/actions/_lib/x`. Dedup is unchanged: ecr computes the hash once via the composite and forwards it to docker-build through the deps-hash input; docker-build computes its own only when the input is empty (the arm-ci direct-call path). Delete the two .sh files — the composite action is now the single shared unit, so no separate script is needed.
The `evict-stale-cache` input on docker-build was inert — no caller set it, so the 14-day deps-tag eviction never ran. The Spark runner is long-lived self-hosted with no ECR, so its local deps-cache tags accumulate indefinitely; it's the legitimate consumer. Set evict-stale-cache: "true" on the arm-ci build so stale tags are pruned each run.
Review fixes (F4-F6): - test_cartpole_training_smoke.py: repoint the perception case from the retired/deprecated Isaac-Cartpole-RGB-Camera-Direct-v0 to the canonical Isaac-Cartpole-Camera-Direct-v0 (defaults to rgb, has an rl_games entry point, not on the removal path); fix the docstring claim about camera envs registering "only" rl_games (they register rl_games + skrl, just no rsl_rl). - isaaclab changelog: add a bullet for the AssetConverterBase default USD dir moving from hardcoded /tmp to tempfile.gettempdir() (a user-facing cross-platform change that was undocumented). - isaaclab_tasks changelog: the package's delta here is test-only (arm_ci markers + aarch64 skip guards in rendering test helpers), so convert the fragment to .skip per AGENTS.md (no user-facing entry). Arm-only test tags: - test_scipy.py / test_torch.py carried [windows_ci, arm_ci]; this is the ARM workflow PR, so scope their tags to arm_ci only. windows_ci test tags belong to the Windows workflow PR.
The develop merge brought in isaac-sim#5891 (split tasks tests into core/contrib), which relocated every test file under test/core/ or test/contrib/ and left only shared helpers + conftest at the test-suite root. The new test_cartpole_training_smoke.py predated that split and was the lone test_*.py still at top-level. Cartpole is a core task, so move it under test/core/ to match the new layout, and bump its _REPO_ROOT path depth from parents[3] to parents[4] (matching the sibling core/ tests).
test_train_cartpole_perception timed out at the default 600s on a cold-cache Spark runner: the first camera-enabled run compiles shaders (~600s) before training even starts, which alone consumes the budget. It passed earlier only because that runner had a warm shader cache. Give the camera case a 1800s timeout (the state case keeps 600s).
Collaborator
Author
Should be fixed. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
arm-cijob inside the existingDocker + Testsworkflow (build.yaml) — sharing itschanges/configgating rather than a standalone workflow. Marker-driven:pytest -m arm_ciselects the aarch64-safe subset. ~28 min cold, ~12 min on a warm runner.pytest.mark.arm_ci; growing the arm canary is a one-line marker, no workflow edit.docker-buildandecr-build-push-pull, so each lives in exactly one place (no drift between the local-store and registry cache checks).1. ARM/Spark CI job
1.1 New
arm-cijob on[self-hosted, arm64]: buildsDockerfile.baseforlinux/arm64via thedocker-buildaction (local cache only — ECR is not wired for arm64, so build and test share one runner), then runs thearm_ci-marked tests viarun-testswithci-marker: arm_ci.1.2 Folded into
build.yamlso the gating (changes,config) and thedocker-build/run-testsactions are shared with the amd64 jobs rather than duplicated in a separate workflow file.1.3 Opts into
docker-build's 14-day stale-cache eviction (evict-stale-cache: true) — the long-lived Spark runner has no ECR, so its localdeps-*tags would otherwise accumulate indefinitely.2. Shared CI utilities
2.1 New
_lib/compute-deps-hashcomposite action — the deps-cache hash (install files + base-image digest). Bothdocker-build(local-store check) andecr-build-push-pull(registry check) call it, so a local hit and a registry hit always resolve the samedeps-<hash>tag. The hash is computed once per build and forwarded to the delegated build via adeps-hashinput.2.2 New
_lib/setup-docker-configcomposite action — credsStore-disabled docker config +nvcr.iologin, idempotent so a caller delegating todocker-builddoes not redo it.2.3
docker-buildis the local-cache build primitive (exact-tag + deps-tag short-circuits);ecr-build-push-pullkeeps the ECR-cache orchestration and delegates the actual build to it.2.4
run-testsgains aci-markerinput, forwarded into the container asCI_MARKERand read bytools/conftest.pyto select test files by marker — the same channel as the existingTEST_FILTER_PATTERN/TEST_INCLUDE_FILESknobs.3. Cross-platform source fixes
3.1
AppLauncherderivesEXP_PATHfrom the installedisaacsimpackage when the env var is unset (the bootstrap early-returns under some aarch64 pip layouts), instead ofKeyError. No-op whenEXP_PATHis already set.3.2
AssetConverterBaseresolves its default USD output undertempfile.gettempdir()instead of a hardcoded/tmp, honoring$TMPDIRand working on Windows.3.3 The kitless rendering test helper skips (rather than fails) the OVRTX/OVPhysX cases on aarch64, where those wheels are not published.
3.4
tools/conftest.pygainsCI_MARKERorchestration: a file-level marker filter so only marker-tagged files spawn per-file subprocesses, parallel to the existingISAACSIM_CI_SHORTpath and not aliased to it.4. Series
Part 2 of
[CI] Cross-platform —. Depends on #5695 (Part 1: marker registration + AppLauncher argv-strip scaffolding). The Windows workflow is #5700 (Part 3).Test plan
arm-cigreen end-to-end on Spark (build + marker-gated tests), cold and warm cache.build.yamljobs unaffected by the composite extraction and thedocker-build/ecr-build-push-pulldelegation.develop(core/contrib test split, Kit-version logging, deferred pxr loading); allarm_ci-marked tests still collected and green after the test-tree move.daily-compatibilityon this branch before merge (gh workflow run "Backwards Compatibility Tests" --ref jichuanh/windows-spark-ci-cartpole).