ci: demo an Allure test report published to GitHub Pages - #3006

Revert this commit before the PR leaves draft.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

run-tests writes one JUnit XML file per suite under test-results/. Each
test job uploads that directory as a test-results-* artifact, and the
final "Check job status" job merges every job's files into two
informational checks:

- "Test Results (aggregate)" (EnricoMi/publish-unit-test-result-action):
  one entry per distinct test across the matrix, with the number of runs
  that failed, plus a PR comment when there are failures.
- "Test Results (by configuration)" (dorny/test-reporter): one section
  per test-results file, so each failure appears under the job
  configuration that produced it.

Both publish steps run with continue-on-error and never fail on test
failures, so the gate in the same job is unchanged. The two reporters
are published side by side for comparison; one of them is meant to be
kept.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Follow-ups from review:

- The nightly PyTorch interop, numba-cuda-mlir, and released cuda-core
  steps in the test workflows still ran pytest with -rxXs. They now use
  -rsxXfE, pipe through the fold script, and write JUnit XML into the
  same test-results/ directory, so every test log ends the same way and
  every suite reaches the "Test Results" checks.
- cuda.bindings gets pytest-github-actions-annotate-failures in its test
  group, so bindings failures are annotated like cuda.core failures. The
  released cuda-core step keeps warning annotations on because its test
  group comes from the released tag, which has no such plugin.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Two problem-matcher files under .github/matchers, each registered by one
step:

- compute-sanitizer.json, registered before the test steps in the Linux
  test workflow: memcheck findings ("Invalid __global__ read ...", the
  non-zero "ERROR SUMMARY" line) become annotations on the job page
  instead of staying buried in the pytest output of sanitizer jobs.
- compiler.json, registered after checkout in the wheel-build workflow:
  gcc/clang "file:line:col: error:", MSVC "file(line): error Cnnnn:",
  and Cython "file.pyx:line:col: message" lines become annotations.
  Errors only, to stay under the per-step annotation cap. Linux builds
  run inside a container, so their paths do not map onto the checkout
  and those annotations stay at the job level.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

One fast row per platform instead of the full matrix, so each CI round
takes about 25 minutes instead of an hour. Revert this commit before the
PR leaves draft.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…y search

dorny/test-reporter with use-actions-summary on writes only to the job
summary and creates no check run, so the "by configuration" report never
reached the PR. EnricoMi's publisher looked the PR up by branch and found
none, because CI runs on copy-pr-bot's pull-request/N mirror of a fork
PR; search_pull_requests finds it by commit and pull_request_build=commit
matches the pushed SHA.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…orters

Third informational publisher in the final job: worded status columns,
a plain-English preamble in the job summary, no annotations, and a PR
comment. Same continue-on-error posture as the other two; one of the
three is meant to be kept.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

… two reporters"

This reverts commit 8483791.

EnricoMi: report_individual_runs off, so a test that fails on every
configuration gets one "All N runs failed" annotation instead of being
split by platform-specific traceback text.

dorny: fail-on-error on, so the per-configuration check is red when any
test failed. The step keeps continue-on-error, so the gate is unaffected.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…omment

Check runs created from inside a workflow are filed under the oldest
workflow that ran on the commit. On a copy-pr-bot mirror that is always
one of the fork PR's own workflows, so both "Test Results" checks showed
up under the assignee/label check. Neither reporter can choose the
suite, so neither creates a check run any more: EnricoMi writes its
counts to the job summary, dorny writes its per-configuration report
there too, and a short "how to read" paragraph precedes both.

The PR comment is now ours: ci/tools/junit_comment.py renders a plain
English summary from the same JUnit files (configurations, tests that
did not pass and where, execution totals, a link to the Summary page)
and marocchino/sticky-pull-request-comment posts it with recreate, so
it reappears at the bottom of the thread on every run.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…ur own comment"

This reverts commit 42c80a7.

Back to check runs for both reporters. The Summary page has no link from
the PR without custom code, and the comment script was more code than the
report is worth. The next commit posts a one-line comment built from the
reporters' own step outputs instead.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Both reporters publish check runs again. The comment on the PR is now a
single line: dorny's passed/failed/skipped counts, and a link to each
check run (EnricoMi's check_url from its json output, dorny's url_html).
EnricoMi's own comment is off so there is one comment, not two.

Also drops list-files from the dorny step, which is not one of its inputs.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

…layout"

This reverts commit 83db5a2.
The injected failures served their purpose on the demo runs.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

This reverts commit 0248098.
The full PR matrix is back.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

# Conflicts:
#	.github/workflows/test-wheel-linux.yml
#	.github/workflows/test-wheel-windows.yml

The final job goes back to main's version: no EnricoMi or dorny check
runs and no test-results comment. The JUnit XML artifacts stay as the
data feed; the report experiments continue in a separate PR.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Revert this commit before the PR leaves draft.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

One fast row per platform. The aarch64 row uses the L4 pool, which had
runners when the A100 pool did not. Revert this commit before any merge.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

The final job renders every test job's JUnit XML as an Allure Report 3
report and publishes it to docs/test-report/pr-N/ on the gh-pages branch,
the way the docs previews are published, with a sticky comment that links
to it. ci/tools/allure_prepare.py tags every suite with its CI
configuration and writes an allurerc.mjs that maps the tags to Allure
environments; without that, Allure folds the same test from every
configuration into retries of one result.

Allure is a CLI from npm run as a plain step, so the actions allowlist
does not apply, and it reads JUnit XML directly. Every step is
continue-on-error; the gate is unchanged.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>