refactor: build two-phase System One progressive code reader

.github/workflows/localization-quality-evaluation.yml

@@ -0,0 +1,223 @@
+name: Localization Quality Evaluation
+
+on:
+  workflow_dispatch:
+    inputs:
+      source_run_id:
+        description: Cross-trace run ID whose localization artifacts should be evaluated
+        required: true
+        default: "35869966400"
+        type: string
+
+permissions:
+  contents: read
+  actions: read
+
+env:
+  SOURCE_RUN_ID: ${{ inputs.source_run_id || '35869966400' }}
+
+jobs:
+  evaluate:
+    name: Blind Claude localization quality review
+    runs-on: ubuntu-24.04
+    environment: ds
+    timeout-minutes: 120
+    env:
+      CLAUDE_MODEL: ${{ vars.CLAUDE_MODEL }}
+      ANTHROPIC_BASE_URL: ${{ vars.ANTHROPIC_BASE_URL || 'https://api.anthropic.com' }}
+    steps:
+      - name: Checkout evaluator code
+        uses: actions/checkout@v4
+        with:
+          path: narness
+          fetch-depth: 1
+          persist-credentials: false
+
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "22.14.0"
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.13"
+
+      - name: Install pinned Claude Code
+        run: |
+          set -euo pipefail
+          npm install --global --no-fund --no-audit @anthropic-ai/claude-code@2.1.278
+          claude --version
+
+      - name: Download frozen localization artifacts
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          set -euo pipefail
+          mkdir -p artifacts/system-one artifacts/claude-reference
+          gh run download "$SOURCE_RUN_ID" \
+            --repo "$GITHUB_REPOSITORY" \
+            --name "code-locator-cross-trace-system-one-$SOURCE_RUN_ID" \
+            --dir artifacts/system-one
+          gh run download "$SOURCE_RUN_ID" \
+            --repo "$GITHUB_REPOSITORY" \
+            --name "code-locator-cross-trace-claude-$SOURCE_RUN_ID" \
+            --dir artifacts/claude-reference
+
+      - name: Resolve frozen subject revision
+        id: source
+        run: |
+          set -euo pipefail
+          python3 - <<'PY' >> "$GITHUB_OUTPUT"
+          import json
+          from pathlib import Path
+          manifest=json.loads(Path("artifacts/system-one/manifest.json").read_text())
+          print(f"repo={manifest['subject_repository']}")
+          print(f"sha={manifest['subject_sha']}")
+          print(f"query={manifest['query']}")
+          PY
+
+      - name: Checkout exact subject revision
+        uses: actions/checkout@v4
+        with:
+          repository: ${{ steps.source.outputs.repo }}
+          ref: ${{ steps.source.outputs.sha }}
+          path: subject
+          fetch-depth: 1
+          persist-credentials: false
+
+      - name: Require evaluator configuration
+        env:
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+        run: |
+          set -euo pipefail
+          test -n "$ANTHROPIC_API_KEY"
+          test -n "$ANTHROPIC_BASE_URL"
+          test -n "$CLAUDE_MODEL"
+
+      - name: Prepare anonymous candidates
+        run: |
+          set -euo pipefail
+          mkdir -p artifacts/quality-evaluation/candidate-a
+          mkdir -p artifacts/quality-evaluation/candidate-b
+
+          python3 narness/research/code-locator/src/localization_quality_evaluation.py prepare \
+            --input artifacts/system-one/localization-result.json \
+            --candidate-id candidate-a \
+            --output-candidate artifacts/quality-evaluation/candidate-a/candidate.json \
+            --output-prompt artifacts/quality-evaluation/candidate-a/prompt.txt
+
+          python3 narness/research/code-locator/src/localization_quality_evaluation.py prepare \
+            --input artifacts/claude-reference/localization-result.json \
+            --candidate-id candidate-b \
+            --output-candidate artifacts/quality-evaluation/candidate-b/candidate.json \
+            --output-prompt artifacts/quality-evaluation/candidate-b/prompt.txt
+
+      - name: Evaluate anonymous candidate A
+        working-directory: subject
+        env:
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+        run: |
+          set -euo pipefail
+          claude -p \
+            --output-format stream-json \
+            --verbose \
+            --model "$CLAUDE_MODEL" \
+            --effort high \
+            --max-turns 80 \
+            --permission-mode dontAsk \
+            --no-session-persistence \
+            --setting-sources '' \
+            --strict-mcp-config \
+            --mcp-config '{"mcpServers":{}}' \
+            --disable-slash-commands \
+            --tools 'Bash,Read,Glob,Grep' \
+            --allowedTools 'Bash(*)' 'Read' 'Glob' 'Grep' \
+            --disallowedTools 'Edit' 'Write' 'Agent' 'Task' 'WebFetch' 'WebSearch' 'EndConversation' 'mcp__*' \
+            --settings '{"disableAllHooks":true}' \
+            < ../artifacts/quality-evaluation/candidate-a/prompt.txt \
+            > ../artifacts/quality-evaluation/candidate-a/evaluator.raw.jsonl
+
+      - name: Evaluate anonymous candidate B
+        working-directory: subject
+        env:
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+        run: |
+          set -euo pipefail
+          claude -p \
+            --output-format stream-json \
+            --verbose \
+            --model "$CLAUDE_MODEL" \
+            --effort high \
+            --max-turns 80 \
+            --permission-mode dontAsk \
+            --no-session-persistence \
+            --setting-sources '' \
+            --strict-mcp-config \
+            --mcp-config '{"mcpServers":{}}' \
+            --disable-slash-commands \
+            --tools 'Bash,Read,Glob,Grep' \
+            --allowedTools 'Bash(*)' 'Read' 'Glob' 'Grep' \
+            --disallowedTools 'Edit' 'Write' 'Agent' 'Task' 'WebFetch' 'WebSearch' 'EndConversation' 'mcp__*' \
+            --settings '{"disableAllHooks":true}' \
+            < ../artifacts/quality-evaluation/candidate-b/prompt.txt \
+            > ../artifacts/quality-evaluation/candidate-b/evaluator.raw.jsonl
+
+      - name: Build scorecards and report
+        run: |
+          set -euo pipefail
+          python3 narness/research/code-locator/src/localization_quality_evaluation.py finalize \
+            --candidate artifacts/quality-evaluation/candidate-a/candidate.json \
+            --raw-jsonl artifacts/quality-evaluation/candidate-a/evaluator.raw.jsonl \
+            --subject-root subject \
+            --model "$CLAUDE_MODEL" \
+            --output-json artifacts/quality-evaluation/candidate-a/scorecard.json \
+            --output-markdown artifacts/quality-evaluation/candidate-a/report.md
+
+          python3 narness/research/code-locator/src/localization_quality_evaluation.py finalize \
+            --candidate artifacts/quality-evaluation/candidate-b/candidate.json \
+            --raw-jsonl artifacts/quality-evaluation/candidate-b/evaluator.raw.jsonl \
+            --subject-root subject \
+            --model "$CLAUDE_MODEL" \
+            --output-json artifacts/quality-evaluation/candidate-b/scorecard.json \
+            --output-markdown artifacts/quality-evaluation/candidate-b/report.md
+
+          python3 narness/research/code-locator/src/localization_quality_evaluation.py report \
+            --candidate-a artifacts/quality-evaluation/candidate-a/scorecard.json \
+            --candidate-b artifacts/quality-evaluation/candidate-b/scorecard.json \
+            --candidate-a-label "System One" \
+            --candidate-b-label "Claude Code" \
+            --output-json artifacts/quality-evaluation/evaluation-report.json \
+            --output-markdown artifacts/quality-evaluation/evaluation-report.md
+
+          python3 - <<'PY'
+          import json
+          import os
+          from pathlib import Path
+          root=Path("artifacts/quality-evaluation")
+          payload={
+              "schema_version": 1,
+              "kind": "code-localization-quality-evaluation-run",
+              "source_run_id": os.environ["SOURCE_RUN_ID"],
+              "subject_repository": "${{ steps.source.outputs.repo }}",
+              "subject_sha": "${{ steps.source.outputs.sha }}",
+              "query": "${{ steps.source.outputs.query }}",
+              "evaluator_model": os.environ["CLAUDE_MODEL"],
+              "candidate_mapping": {
+                  "candidate-a": "System One",
+                  "candidate-b": "Claude Code",
+              },
+          }
+          (root/"manifest.json").write_text(
+              json.dumps(payload, indent=2, ensure_ascii=False)+"\n"
+          )
+          PY
+
+          cat artifacts/quality-evaluation/evaluation-report.md >> "$GITHUB_STEP_SUMMARY"
+
+      - name: Upload quality evaluation
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: code-locator-quality-evaluation-${{ github.run_id }}
+          path: artifacts/quality-evaluation/
+          if-no-files-found: warn
+          retention-days: 90