voltry-probe 0.2.0__tar.gz → 0.2.2__tar.gz

{voltry_probe-0.2.0 → voltry_probe-0.2.2}/.gitignore

@@ -48,3 +48,16 @@ next-env.d.ts
 
 # hypothesis example database
 .hypothesis/
+
+# --- Field-capture run artifacts (tools/capture) ---
+# Raw rental artifacts (logs, sidecars, ssh session records) stay local; verified
+# bundles are published deliberately into apps/console/fixtures/captures/.
+captures/runs/
+
+# --- Census index + snapshots (tools/census) ---
+# Marketplace inventory is large, host-identifying, and regenerated; keep it local.
+captures/census/
+
+# --- Local Phase-0 registry (tools/registry/seed_local.py) ---
+# Seeded internal registry (SQLite DB + object store); regenerated from the field bundles.
+captures/registry/

{voltry_probe-0.2.0 → voltry_probe-0.2.2}/CHANGELOG.md

@@ -3,6 +3,60 @@
 All notable changes to `voltry-probe` are documented here. The format follows
 Keep a Changelog.
 
+## [0.2.2] - 2026-07-05
+
+Measured-honesty fix in the NVLink reader, found by independent audit.
+
+- NVLink error counters are now banked per counter, and only when actually
+  measured: the counter's NVML_NVLINK_ERROR_DL_* constant must resolve and at
+  least one per-link read must succeed. 0.2.1 guarded only the all-or-nothing
+  case; on current nvidia-ml-py (12.575.51, 13.610.43), which dropped
+  NVML_NVLINK_ERROR_DL_CRC but kept REPLAY/RECOVERY and the counter API, it
+  banked a fabricated crc_errors: 0 even though CRC was never read. The key is
+  now omitted (renders Not Assessed), matching how the probe treats all other
+  absent telemetry. Same treatment when the constants resolve but the device
+  refuses every per-link counter read.
+- A genuinely-measured zero still banks as 0: a clean read is a valid fact.
+
+## [0.2.1] - 2026-07-05
+
+Hardening for the live path on the full range of rental and consumer hardware,
+based on the NVML per-device-class support matrix. Previously a live scan of a
+device outside the certifiable envelope died with a raw driver traceback.
+
+- Honest refusals: a live scan of a device that cannot expose the required
+  memory-health reads now raises `UnsupportedGpuError` (exported at package
+  level) with a diagnosis naming the device and the reason, and `voltry scan`
+  exits with a clean, specific error instead of a traceback. The cases:
+  - consumer GPUs with no ECC memory (GeForce class), disambiguated from
+  - ECC-capable devices with ECC currently disabled (common on workstation
+    cards such as the RTX A6000, which ship ECC-off), with the fix named;
+  - pre-Ampere devices with no row remapping (Tesla V100, T4, P100 class);
+  - MIG-enabled devices, where row remapping is unreadable even on the parent
+    handle.
+- Board serial is now an optional read: on devices with no inforom serial
+  (GeForce class, some virtualized environments) the reader omits the field
+  instead of crashing; identity binds to the GPU UUID and the bundle records
+  serial "UNKNOWN". Never fabricated.
+- A generic NVML error on a required read still propagates unchanged: a driver
+  fault stays a driver fault, distinct from an honest device-class refusal.
+- More ratchet fields sealed on every successful scan: MIG state (closing the
+  gap where MIG appeared only in refusal messages), GPU core count, HBM memory
+  temperature via the field-values API (banked only when the per-field status
+  is clean; GDDR cards without the sensor omit it), and the per-bank
+  row-remapper histogram where the binding exposes it.
+- NVLink capture survives current nvidia-ml-py builds, which dropped the legacy
+  error-counter API (the NVML_NVLINK_ERROR_DL_* constants and
+  nvmlDeviceGetNvLinkErrorCounter). Link state is still captured; the error
+  counters are omitted when the binding lacks the API, rendering Not Assessed
+  instead of crashing the scan. Found live on a rented A100 SXM4, where 0.2.0
+  dies with an AttributeError on any NVLink-bearing card.
+- Refusal diagnoses only assert what was actually read: the pre-Ampere claim is
+  made only when the device architecture was read and predates Ampere, and the
+  no-ECC claim only on a clean NotSupported from the ECC-mode read (virtualized
+  guests that hide board-level reads get an honest "not readable here" message
+  instead of a fabricated device-class fact).
+
 ## [0.2.0] - 2026-07-04
 
 - Lifetime duty (the odometer) on the certificate: GPU-hours, thermal cycles,

{voltry_probe-0.2.0 → voltry_probe-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: voltry-probe
-Version: 0.2.0
+Version: 0.2.2
 Summary: Voltry Probe — the open, read-only agent. Reads NVML/DCGM/Redfish/attestation into a signed evidence bundle and renders the offline certificate.
 License-Expression: Apache-2.0
 License-File: LICENSE

{voltry_probe-0.2.0 → voltry_probe-0.2.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "voltry-probe"
-version = "0.2.0"
+version = "0.2.2"
 description = "Voltry Probe — the open, read-only agent. Reads NVML/DCGM/Redfish/attestation into a signed evidence bundle and renders the offline certificate."
 requires-python = ">=3.10"
 license = "Apache-2.0"

{voltry_probe-0.2.0 → voltry_probe-0.2.2}/tests/test_cli.py

@@ -89,6 +89,25 @@ def test_scan_then_cert_offline(tmp_path):
     assert "UNVERIFIED" not in html  # a signed, valid bundle renders as verified
 
 
+def test_scan_live_refusal_is_clean_and_specific(monkeypatch):
+    # A live scan of a device outside the certifiable envelope (no ECC, pre-Ampere, MIG)
+    # must exit nonzero with the diagnostic reason, never a raw driver traceback.
+    from voltry_probe.sources import UnsupportedGpuError
+    from voltry_probe.sources.live import LiveSource
+
+    def _refuse(self):
+        raise UnsupportedGpuError(
+            "NVIDIA GeForce RTX 4090: this is a consumer-class GPU with no ECC memory."
+        )
+
+    monkeypatch.setattr(LiveSource, "capture", _refuse)
+    result = runner.invoke(app, ["scan", "--ephemeral-key"])
+    assert result.exit_code == 2
+    text = _text(result)
+    assert "no ECC" in text
+    assert "Traceback" not in text
+
+
 def test_scan_requires_a_signing_key():
     # No --signing-key and no --ephemeral-key → refuse (don't silently sign).
     result = runner.invoke(app, ["scan", "--fixture", str(FIXTURE)])