ci(pr): classify PR test failures as new vs. known and post sticky comment

[rgsl888prabhu]

Summary

• For every PR test job, cross-reference each failure against the target branch's nightly failure history and post a single sticky comment that splits failures into NEW (introduced by this PR) vs. KNOWN (recurring on nightly, known flaky on nightly, or flaked in this run via pytest-rerunfailures).
• No changes to retry behavior; reuses the existing nightly_report.py JUnit parser and history schema.
• Sticky single comment per PR, identified by hidden marker <!-- pr-test-classification -->, updated in place on each push. Skipped entirely when nothing failed or flaked.

How it works

1. Per-matrix classification (PR mode). ci/utils/nightly_report_helper.sh now branches on RAPIDS_BUILD_TYPE. For PR runs it points --s3-history-uri at the target branch's nightly history (read-only — never writes back) and uploads a per-matrix summary to a PR-scoped prefix s3://.../ci_test_reports/pr/run-${GITHUB_RUN_ID}/.
2. Read-only classifier. nightly_report.py --mode pr annotates each failure with pr_classification:
   • new — not in nightly history, or only there as resolved and not flagged flaky.
   • known_recurring — nightly history status=active.
   • known_flaky_nightly — nightly history is_flaky=true (regardless of status).
   • known_flaky_pr — passed on retry within this run, not previously known to flake.
3. Aggregator. ci/utils/aggregate_pr.py lists the run-scoped prefix, merges per-matrix summaries (via shared helpers in the new aggregate_common.py), and produces a Markdown comment with two top-level sections (NEW, KNOWN) and three sub-groups inside KNOWN.
4. Comment poster. ci/pr_summary.sh looks up an existing comment by its hidden marker and either PATCHes it in place or POSTs a new one.
5. Workflow wiring. A new pr-test-summary job in .github/workflows/pr.yaml runs after every PR test job (if: always(), permissions: pull-requests: write). It is not added to pr-builder's needs: list — comment posting is informational and must not block merging.

Files

File	Purpose
ci/utils/aggregate_common.py (new)	Shared download_summaries, load_local_summaries, aggregate_summaries, html_escape
ci/utils/aggregate_nightly.py	Imports from aggregate_common; behavior unchanged
ci/utils/aggregate_pr.py (new)	Builds the Markdown PR comment body
ci/utils/nightly_report.py	New --mode pr; preserves pr_classification in summary JSON
ci/utils/nightly_report_helper.sh	PR-mode S3 wiring (read target-branch history; write run-scoped summary)
ci/pr_summary.sh (new)	Aggregates and posts/updates the sticky comment
.github/workflows/pr.yaml	New pr-test-summary job

Test plan

• Trigger CI on this PR and confirm pr-test-summary runs after the test jobs and produces (or updates) a sticky comment when there are failures.
• Force a known-recurring failure in the test matrix; confirm it lands in the KNOWN — Already broken on nightly section.
• Force a brand-new failure (rename a passing test to fail); confirm it lands in NEW failures with classification new.
• Confirm a flaky-passed-on-retry test (already supported by pytest-rerunfailures) lands in KNOWN — Flaked in this PR run.
• Confirm an all-green PR run does not post a comment.
• Confirm re-running the workflow updates the existing comment in place rather than creating a new one.
• Confirm nightly runs are unaffected: history file still updates, consolidated dashboard renders, Slack notification still fires.
• Confirm ${{ github.token }} with permissions: pull-requests: write is sufficient (the only existing token usage in pr.yaml is read-only via gh pr view); fall back to a PAT if not.
• Verify CUOPT_AWS_* credentials have write access to the new ci_test_reports/pr/run-*/ prefix.

Follow-ups (not in this PR)

• S3 bucket lifecycle rule to expire ci_test_reports/pr/run-*/ after ~14 days.

🤖 Generated with Claude Code

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[rgsl888prabhu]

/ok to test c4876d2