[Tool] unstable-case workflows: wire recovered-pool / drip (probation v2)

Companion to StarRocks/ci-tool probation-flow-v2 (§13.1/§13.2). Must merge
together with it.

- unstable-case-review.yml: report now passes --recovered-pool (discovery-only,
  drops --probation/--batch); the PR step commits unstable_case_recovered_pool.json
  and is retitled "update recovered pool" (review no longer grants probation).
- unstable-case-promote-demote.yml: passes --recovered-pool + --target-probation
  (new optional input, default 30); git add includes the pool file so drip/pool
  maintenance is committed. The existing `git diff --quiet files/` gate already
  detects pool changes.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: huailiu1122 <liuhuaiyuan@starrocks.com>

Author: huailiu1122

.github/workflows/unstable-case-promote-demote.yml

@@ -31,6 +31,11 @@ on:
         required: false
         default: '2'
         type: string
+      target_probation:
+        description: 'PROBATION target watermark — drip fills probation up to this many cases from the recovered pool'
+        required: false
+        default: '30'
+        type: string
       dry_run:
         description: 'Only print decisions; do not write files or open a PR'
         required: false
@@ -73,6 +78,8 @@ jobs:
             --resolve-record files/unstable_case_resolve_record.json \
             --repo "${{ github.repository }}" --branch "${BRANCH}" \
             --k "${{ inputs.k || '5' }}" --min-days "${{ inputs.min_days || '2' }}" --q "${{ inputs.q || '2' }}" \
+            --recovered-pool files/unstable_case_recovered_pool.json \
+            --target-probation "${{ inputs.target_probation || '30' }}" \
             ${DRY}
 
           if [[ -n "${DRY}" ]]; then
@@ -100,7 +107,7 @@ jobs:
           # push is rejected non-fast-forward — making the workflow un-retriable.
           BR="auto/probation-promote-demote-$(date +%Y-%m-%d)-${ch_branch}-r${GITHUB_RUN_ID}.${GITHUB_RUN_ATTEMPT:-1}"
           git checkout -b "$BR"
-          git add files/unstable_case.json files/unstable_case_probation.json files/unstable_case_details.json files/unstable_case_resolve_record.json
+          git add files/unstable_case.json files/unstable_case_probation.json files/unstable_case_details.json files/unstable_case_resolve_record.json files/unstable_case_recovered_pool.json
           git commit -m "[Tool] probation promote/demote ${BRANCH} (auto)"
           git push origin "$BR"
           BODY=$(mktemp)

.github/workflows/unstable-case-review.yml

@@ -275,8 +275,7 @@ jobs:
             --be-oss-path  "${BE_OSS}"  \
             --sql-oss-path "${SQL_OSS}" \
             --output-dir "${WORK_DIR}" \
-            --probation files/unstable_case_probation.json \
-            --batch 30
+            --recovered-pool files/unstable_case_recovered_pool.json
 
           echo "WORK_DIR=${WORK_DIR}" >> $GITHUB_OUTPUT
 
@@ -313,13 +312,12 @@ jobs:
           if-no-files-found: warn
           retention-days: 14
 
-      - name: Open PR granting probation eligibility
-        # Opt-in via the grant_probation input (default false = off, same as
-        # before). Needs SYNC_PAT (configured on CelerData) and the ubuntu/ai
-        # runners (CelerData only), so in practice this only fires on CelerData.
-        # The earlier blockers are resolved: SYNC_PAT is now set, and the
-        # runner's /var/lib/ci-tool is a bind-mount real dir so the per-job
-        # `cp` is a true isolated copy.
+      - name: Open PR updating recovered pool
+        # §13.1/§13.2: review is discovery-only — it writes recovered candidates to
+        # unstable_case_recovered_pool.json; the promote-demote workflow drips them
+        # into probation. This step opens a PR carrying the pool update.
+        # Opt-in via the grant_probation input (kept for backward compat — same
+        # cron/dispatch flag). Needs SYNC_PAT (CelerData) and ubuntu/ai runners.
         if: ${{ inputs.grant_probation == 'true' }}
         id: open_pr
         env:
@@ -329,9 +327,9 @@ jobs:
           git config user.name  "wanpengfei-git"
           git config user.email "wanpengfei91@163.com"
 
-          # Grant-eligibility touches these three files; skip if none changed.
-          if git diff --quiet files/unstable_case.json files/unstable_case_probation.json files/unstable_case_details.json; then
-            echo "No probation-eligibility changes this run — skipping PR"
+          # Discovery-only now touches just the recovered pool; skip if unchanged.
+          if git diff --quiet files/unstable_case_recovered_pool.json; then
+            echo "No recovered-pool changes this run — skipping PR"
             echo "pr_url=" >> $GITHUB_OUTPUT
             exit 0
           fi
@@ -340,10 +338,10 @@ jobs:
           # RUN_ATTEMPT increments), so include RUN_ATTEMPT or a retry after a
           # partial failure reuses the branch and the push is rejected
           # non-fast-forward.
-          BR="auto/probation-eligibility-$(date +%Y-%m-%d)-${COMMIT_SHA:0:8}-r${GITHUB_RUN_ID}.${GITHUB_RUN_ATTEMPT:-1}"
+          BR="auto/recovered-pool-$(date +%Y-%m-%d)-${COMMIT_SHA:0:8}-r${GITHUB_RUN_ID}.${GITHUB_RUN_ATTEMPT:-1}"
           git checkout -b "$BR"
-          git add files/unstable_case.json files/unstable_case_probation.json files/unstable_case_details.json
-          git commit -m "[Tool] grant probation eligibility ${BRANCH} ${COMMIT_SHA:0:8} (auto)"
+          git add files/unstable_case_recovered_pool.json
+          git commit -m "[Tool] update recovered pool ${BRANCH} ${COMMIT_SHA:0:8} (auto)"
           git push origin "$BR"
 
           # YAML's run: | block strips the common indent, so the heredoc body and EOF
@@ -352,16 +350,16 @@ jobs:
           cat > "$BODY_FILE" <<EOF
           Auto-generated by [unstable-case-review run #${GITHUB_RUN_ID}](https://github.com/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}).
 
-          Cases that passed all 3 isolation rounds are **moved into probation** (not removed): they leave \`unstable_case.json\`, enter \`unstable_case_probation.json\` (stamped \`probation_since\`), and must now earn promotion from real PR-concurrency evidence. An isolated green is not PR-concurrency stable, so this only grants a trial — promotion/demotion is decided by the promote-demote workflow.
+          Cases that passed all 3 isolation rounds are added to the **recovered pool** (\`unstable_case_recovered_pool.json\`). Review is now discovery-only: the pool is the durable candidate list, and the **promote-demote workflow drips** cases from it into probation up to the target watermark (no re-running this regression per batch). An isolated green is not PR-concurrency stable, so probation + promote-demote still decide the final outcome.
 
-          Please sanity-check the diff — cases with a high \`add_count\` (yo-yo counter) are riskier.
+          Please sanity-check the diff.
           EOF
 
           PR_URL=$(gh pr create \
             --repo StarRocks/ci-tool \
             --base main \
             --head "$BR" \
-            --title "[Tool] grant probation eligibility $(date +%Y-%m-%d) (auto)" \
+            --title "[Tool] update recovered pool $(date +%Y-%m-%d) (auto)" \
             --body-file "$BODY_FILE")
           rm -f "$BODY_FILE"
           echo "pr_url=${PR_URL}" >> $GITHUB_OUTPUT