feat(results): materialize per-test task bundles

[christso]

Summary

AgentV runs can now carry a native, per-test task bundle beside each result row. Instead of a parallel run-source schema, each completed result can point from index.jsonl to task/EVAL.yaml, task/targets.yaml, copied input files, and copied grader assets, giving future audit/rerun flows a normal AgentV eval contract to consume.

This hard-deprecates the unreleased run-source.json surface rather than preserving a compatibility reader. Historical runs still load without task metadata; new bundle-capable runs expose focused index links (artifact_dir, task_dir, eval_path, targets_path, files_path, graders_path) and Dashboard file APIs can traverse those task paths.

Key design choices:

Area	Decision
Portable unit	Per-test task/ folder, not run-level source/recipe JSON
Rerunnable source	Native EVAL.yaml plus selected targets.yaml
Secret handling	Preserve ${{ ENV_VAR }} placeholders; redact literal secret-looking values and secret files
Output safety	Materialization never creates .agentv/results under the test artifact folder or task/
Dry-run aliases	Result rows may use resolved target names, while task/EVAL.yaml preserves the selected target name

Verification

• bun test apps/cli/test/commands/eval/task-bundle.test.ts apps/cli/test/commands/eval/artifact-writer.test.ts apps/cli/test/commands/results/serve.test.ts apps/cli/test/commands/results/combine.test.ts (128 pass)
• bun --filter agentv typecheck
• Pre-push hook: bun --filter @agentv/core typecheck && bun --filter @agentv/phoenix-adapter typecheck && bun --filter agentv typecheck, then biome check .

Red/Green UAT

Used the same dry-run eval with input_files on origin/main and this branch. The dry-run quality score was the same on both sides; the checked behavior was artifact materialization.

State	Evidence
Red (origin/main)	index.jsonl had no task_dir; no task/EVAL.yaml; no copied task file (/tmp/agentv-red-uat.nlqOYC)
Green (this branch)	index.jsonl included task_dir, eval_path, targets_path, and files_path; task/EVAL.yaml preserved target: mock-target; the fixture was copied under task/files/external/...; no nested .agentv/results (/tmp/agentv-green-uat.SKAgFt)

---