[Burntgogi]

Addresses the container drive-device item of #524; does not close the issue's separate class-target item.

Fix

• One drive_source owner uses findmnt --nofsroot; drive_dev resolves PKNAME, then KNAME for a whole-disk alias, then retains the existing basename/counter guard.
• engram-rate uses the shared drive reader instead of duplicating discovery; nvtier's witness mount-source lookup uses the same source owner. The existing depth/nvtier drive_dev callers inherit the fix.
• Missing mount/counters remains a named exit 2, before timing. The Spark container documentation states the required readable /sys/block/<disk>/stat.

Verification

• FAIL-first existing card harness reproduced /sys/block/models]/stat for /dev/testp1[/models] and the missing-KNAME alias failure.
• After the fix LC_ALL=C bash tools/check-recipes.sh --block cardtests: 874 tests, 0 failed. The five new cases cover bind roots, whole-disk aliases, canonical fallback, unreadable counters and unreadable mounts, without requiring Linux /sys or a real disk.
• check-recipes.sh --block fnowner: 191 scripts, 9 fixtures, passed.
• check-recipes.sh --block recipescheck: 149 gate rows held to derivation, passed.
• nvtier-read.sh --self-test: 7 checks, 0 failures; --parse-only --rounds 1 passed; bash -n on all changed shell scripts passed.
• Real, untimed workstation smoke: drive_source Cargo.toml → /dev/nvme0n1p2, drive_dev Cargo.toml → nvme0n1, drive_sectors → numeric kernel counter. This host's mount is not the reporter's DGX bind mount; the bind behavior is the deterministic FAIL-first regression. No DGX/NVIDIA timed arm or throughput result is claimed.

[github-actions]

Verdict: CLEAN: The shared reader resolves bind-root sources and whole-disk aliases while preserving named refusals before timing.

Checked

Duplicates: 5 new functions checked, including test stubs; no duplicate owners found. Mount-source discovery has one production owner.

Complexity: no Rust functions changed. The production change adds a source resolver and one guarded device fallback.

Silent failure: both suppressed lsblk results lead to the readable-counter guard. Mount and counter failures reach caller exit 2.

Tests: all 5 isolated drive cases and syntax checks on all 4 changed shell files passed. Removing --nofsroot or the mount diagnostic at drive-read.sh:11, KNAME at drive-read.sh:21, basename at drive-read.sh:22, or the counter guard at drive-read.sh:23 made a corresponding assertion fail. Successful but empty PKNAME output and a spaced target argument also passed.

Gates: no thresholds moved or existing cases removed. The PR reports FAIL-first reproduction for bind roots and aliases.

Escape catalog: no C1–C4 defects confirmed. The three runners share discovery, and named diagnostics expose refusal paths.

Notes

The full card harness and real hardware were not run during validation.

_{🔨 Shingler, bloomery's review bot: approves a clean pull request and requests changes on a must-fix (how it works)}

[juliankang4]

@Burntgogi Thanks for the contribution. The review is done, so I'm merging this.