Fix reusable artifact validation for nested metadata / 修复嵌套元数据的可复用产物校验

[cquil11]

Summary

• canonicalize nested JSON values before using them in reusable-artifact identities
• preserve KV offload backend name/version/options while making mappings and lists hashable
• add regression tests for the exact {"name": "native"} failure and deeper nested metadata
• verify duplicate detection still reports duplicates instead of crashing

Root cause: post-merge run 29425167775 passed a dict-valued kv_offload_backend into Counter, which raised TypeError: unhashable type: 'dict' while reusing PR #2158's staged artifacts.

Validation

• python -m pytest utils/test_validate_reusable_sweep_artifacts.py -v — 26 passed
• complete test-changelog-gate.yml pytest command — 140 passed
• replayed the exact failing artifact from run 29359203708; its identity is now hashable

中文说明

• 在构造可复用产物标识前，递归规范化嵌套 JSON 值
• 保留 KV 卸载后端的名称、版本和选项，同时将映射与列表转换为可哈希形式
• 为本次 {"name": "native"} 故障以及更深层的嵌套元数据添加回归测试
• 验证重复产物检测会正确报告重复项，而不会崩溃

根因：合并后运行 29425167775 在复用 PR #2158 的暂存产物时，将字典类型的 kv_offload_backend 传入 Counter，触发 TypeError: unhashable type: 'dict'。

验证

• python -m pytest utils/test_validate_reusable_sweep_artifacts.py -v — 26 项通过
• 完整执行 test-changelog-gate.yml 中的 pytest 命令 — 140 项通过
• 使用运行 29359203708 中触发故障的原始产物复测；其标识现在可正常哈希

[claude]

Claude finished @cquil11's task in 1m 50s —— View job

---

Review of PR #2221

• Gather context (diff, modified files)
• Review canonicalization logic and tests
• Post review feedback

LGTM - no blocking issues found

The freeze_identity_value canonicalization is correct: JSON object keys are always strings, so the sorted() over (key, frozen_value) pairs is safe (unique keys mean comparison never falls through to the values), and preserving list order is the right call for identity purposes — the regression test confirms ["cpu"] and ["cpu", "gpu"] produce distinct identities while key order is normalized. Both agentic point rows and aggregate rows flow through the same agentic_key, so the frozen identities stay consistent across the duplicate check (Counter) and the point-vs-aggregate set comparison. Test coverage for the original {"name": "native"} crash, nested metadata, and end-to-end duplicate detection is solid.

Note: I could not run pytest in this environment (command approval denied), but the PR's CI and the author-reported 26/26 + 140/140 passing runs cover that.

[claude]

LGTM — a minimal, well-tested fix for the unhashable-dict crash in artifact identity building.

Extended reasoning...

Overview

The PR adds a single helper, freeze_identity_value, to utils/validate_reusable_sweep_artifacts.py, and applies it at the one call site (agentic_key) where a dict-valued kv_offload_backend field was being fed unmodified into a tuple that later goes through Counter/set operations, causing TypeError: unhashable type: 'dict'. The helper recursively converts dicts into a sorted tuple of (key, value) pairs (order-independent) and lists/tuples into tuples, making arbitrarily nested JSON metadata hashable and comparable. Test-only changes add two regression tests: one asserting the identity is hashable and order-independent across nested dict/list values, and one exercising the full validate_agentic_artifacts duplicate-detection path with a mapping-valued kv_offload_backend.

Security risks

None. This is internal CI validation tooling for benchmark artifact reuse; it does not process untrusted input, execute code, or touch auth/permissions.

Level of scrutiny

Low-to-moderate. This is a narrowly scoped bug fix in a non-production, CI-support utility (artifact validation for benchmark result reuse), not a critical runtime path. The change is small, self-contained, and directly matches the stated root cause (a real CI failure from run 29425167775).

Other factors

I ran the full test suite (utils/test_validate_reusable_sweep_artifacts.py) locally and confirmed all 26 tests pass, matching the PR's validation claims. I also traced kv_offload_backend back to utils/agentic/aggregation/process_agentic_result.py and utils/matrix_logic/generate_sweep_configs.py, confirming it is indeed dict-shaped (name/version/options) in the real pipeline, so the fix addresses a genuine, reachable bug rather than a hypothetical one. The recursive freeze logic is straightforward and correctly handles nesting (dicts of lists of dicts, etc.) via sorted-tuple canonicalization, with no edge cases (e.g., non-string dict keys) that would arise from real JSON metadata.