os_update: address GPT partitions by PARTLABEL, use layout offsets only without one

[mobileoverlord]

avocado deploy to an imx8mp-evk failed in os_update:

Writing rootfs_hash -> /dev/mmcblk1@618659840 (partition: rootfs-b-hash)
Failed to open device /dev/mmcblk1 for rootfs_hash: No such file or directory

What

When the bundle carries a layout (stone always emits one), artifacts were written by raw offset to the layout's device path even though /dev/disk/by-partlabel/<name> existed. The manifest's devpath is kernel-dependent (the eMMC is mmcblk1 in the manifest, mmcblk2 under 6.18), and computed offsets need not match where fwup placed the partitions.

How

locate_target resolves the PARTLABEL first. The layout offset is used only when the disk carries no labels at all (MBR layouts). A labeled disk that lacks this one label is a layout mismatch — a device provisioned before the partition existed — and is refused with a message saying to reprovision, rather than written at a computed offset onto whatever is there now. Both the staged and the streaming write paths use it.

Tests: unlabeled + layout → offset; unlabeled without layout → error.

https://claude.ai/code/session_01S75qGgqVA2cQUsZdW46qUz

[mobileoverlord]

Follow-up (stone has issues disabled, noting here): stone bundle always embeds layout, so an avocadoctl older than this PR still takes the offset path on GPT machines. Emitting layout only for label-less/MBR layouts would let old readers fall back to PARTLABEL on their own; manifests carry no GPT/MBR marker today, so that needs a manifest-level signal.

[Copilot]

Pull request overview

This PR fixes os_update device targeting by preferring kernel-resolved GPT PARTLABEL paths (/dev/disk/by-partlabel/...) over bundle layout device paths/offsets, and only using layout offsets as a fallback when PARTLABELs aren’t available.

Changes:

• Introduces locate_target/WriteTarget to route writes to either a resolved partition path (preferred) or a raw device + byte offset (fallback).
• Applies the same targeting logic to both staged (apply_os_update) and streaming (apply_os_update_streaming) write paths.
• Adds unit tests covering the “no PARTLABEL + layout -> offset” and “no PARTLABEL + no layout -> error” behaviors.

Suppressed comments (1)

src/os_update.rs:1576

• Same flakiness concern as the test above: use a deliberately-unique PARTLABEL so the test result doesn’t depend on the host’s attached disks.

    #[test]
    fn an_unlabeled_partition_without_a_layout_is_an_error() {
        assert!(locate_target("avocadoctl-test-no-such-label", None).is_err());
    }

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

Suppressed comments (1)

src/os_update.rs:594

• Path::exists() follows symlinks and returns false on metadata errors, so a dangling PARTLABEL or an EACCES/I/O failure still reaches the raw-offset fallback. This contradicts the fail-safe behavior described above and can write to the wrong location. Inspect the directory entry without following it, and only treat a definite NotFound as absence.

    if label.exists() {