test: pin platform_fingerprint degrade paths (PRPUNDIT-12)

[jiagaoxiang]

Summary

• Exercises platform_fingerprint three-tier degrade: probe None → unavailable, outer exception → status=error, and per-block gpu/stack status=error.

Closes test gap PRPUNDIT-12.

Test plan

• PYTHONPATH=src pytest src/hyperloom/common/tests/test_platform_probe.py -k platform_fingerprint

[zoroyihan7]

Thanks — all three degrade tiers are reached and the assertions are on the real returned dict (including the exact "no host CPU sysfs on this machine" literal), no mock-asserts-the-mock anywhere.

Two things mutation testing turned up.

1. Per-block independence isn't actually pinned. test_platform_fingerprint_gpu_and_stack_blocks_degrade breaks both amdgpu_device_count and detect_stack_fingerprint in the same call, so it can never observe a healthy block surviving its sibling's failure — which is the "the rest of the record survives" half of the contract. Mutating the stack handler to also overwrite record["gpu"] and record["cpu"] leaves all 15 tests green. Worse, swapping which key each handler writes (GPU handler setting record["stack"] and vice versa) is undetectable, and in the single-block-failure case silently drops the "gpu" key entirely.

Splitting into two cases — gpu raises / stack healthy, and stack raises / gpu healthy — and asserting the surviving block's real content in each (e.g. got["stack"] != {"status": "error"}, got["gpu"]["gfx_arch"] present) closes both.

2. The outer net is pinned for only one exception type. Narrowing except Exception as exc to except RuntimeError as exc stays green, because the new test happens to raise RuntimeError. An OSError/AttributeError escaping probe_cpu_platform would then propagate to the crash-path caller with no test noticing — and per exporter.py:1010-1013 there's no second net there, so reports/final.json would be lost for a run that already died. Raising a non-RuntimeError in one of the cases fixes it.

Smaller ones:

• "multi_node_session": multi_node → bool(multi_node) stays green, which destroys the docstring's explicit None-vs-False distinction ("None records that nobody established it ... rather than an unearned False") — no test calls platform_fingerprint with multi_node left unset on an ok record.
• No test covers the all-healthy status="ok" record, so the gpu sub-dict shape (host_count / gfx_arch / amdgpu_driver, the "unknown" fallbacks) has no pin — deleting or renaming any of those keys keeps the suite green. Deleting the "host": socket.gethostname() line is green too.
• The new tests monkeypatch module globals, which the module docstring (lines 24-25) discourages in favour of the injectable root. It's justified here — platform_fingerprint takes no root and the error tiers aren't reachable through a fake sysfs tree — but a one-line comment saying so would stop the next reader "fixing" it.

[jiagaoxiang]

Split GPU-raises/stack-healthy vs stack-raises/GPU-healthy (asserting the surviving block's real content), and the outer net now raises OSError. Added the all-healthy status=ok gpu sub-dict (host_count / gfx_arch / amdgpu_driver, including "unknown" fallbacks), host=gethostname(), and multi_node is None vs False. Comment on the tests explains why these cases monkeypatch module globals (platform_fingerprint has no injectable root).

[zoroyihan7]

Re-reviewed. Splitting into two single-block-failure cases is exactly what was needed — the surviving-block content assertions are what make the key-swap mutations detectable.

Five mutations, all red now:

• except Exception → except RuntimeError on the outer net (L263) — red
• gpu handler writing record["stack"] instead of record["gpu"] (L256) — red
• stack handler writing record["gpu"] instead of record["stack"] (L261) — red
• multi_node → bool(multi_node) (L236) — red
• host key renamed (L235) — red

17 passed. LGTM.