RTK-SLAM acquisition hardening — 2026-08-17

Implementation: 25d712b7, capacity correction 2a997cb9, and attached storage recovery 0c3f5884

Outcome: BLOCKED_INSUFFICIENT_SPACE / ACTIONABLE

Benchmark profiles satisfied by this work: 0

The two RTK-SLAM release profiles remain NO_DATA, but obtaining their public inputs no longer starts an unbounded multi-gigabyte download. The acquisition tool now builds a source-pinned, resume-aware capacity plan before any network or write side effect, and gives one copy-ready retry for a larger filesystem.

Immutable source contract

The dataset is RTK-SLAM Dataset, CC-BY-4.0, fixed at revision 87619d2da3f345109b9a2b0d3a192a8596b4d2d3. The official tree API supplied the exact DB3 byte counts and LFS SHA-256 identities. The small metadata files were hashed through the same immutable revision URLs.

Sequence	DB3 bytes	DB3 SHA-256	Metadata bytes	Metadata SHA-256
Construction Seq2	10,656,112,640	`9e808703a57d7be6afa6a37abb8f5d65c6566f71f4864cd4c24cb01f6ab82af5`	1,930	`2cc6cb1e4a53b2d1c371499e489582b88737b1dbce79f01df4b6c811e43db8ff`
Construction Seq1	13,180,936,192	`adf7e5e8f8d73a0a3a0c09f80d846ca3a88446809ee391a818260d4bd3d03a7a`	1,930	`ad9b7f8b01862305740abd83e671dbc1c080c8d6e2b82a9e05b1152306382318`
Stadtgarten Seq2	16,793,665,536	`d303eeaa773ae1606ddbefb38c509009d9ef9c81a7d4bc6b68869db4287d27e2`	1,928	`9538fe9ca6b3d496bd35eb13519e38136a4a86c2a857f1f0215a2ded6a62037c`
Stadtgarten Seq1	30,263,574,528	`6f674fff7182e54d2aa12cac36b0be36d022f67e8624b3df4d1acf8018fa6e5b`	1,933	`b1fa518aead0436fd574db48b3425d3f1fdcb4a9e26a752f1be0733ed9af3aae`

Surveyed checkpoints and example trajectories come from the official RTK-SLAM evaluation repository, fixed at detached commit f2921a58caf5a87c1f4f73b48c6f2a5e35f92924. The tool initializes an empty repository and fetches that exact commit; it does not shallow-clone a moving default branch. Reuse requires the exact clean commit and all four ground-truth CSV files.

Capacity result on the implementation carrier

This command was run from a clean 2a997cb9 checkout:

python3 scripts/download_rtk_slam_dataset.py \
  --sequence construction_seq2 \
  --eval-assets \
  --dest /tmp/rtk-slam-live-block-evidence

The normal, non-dry-run request exited 2 before creating the destination or starting wget/Git network work:

Field	Exact bytes
remaining payload, including 150 MB eval allowance	10,806,114,570
10% filesystem reserve	1,080,611,457
required additional working space	11,886,726,027
observed free space	6,409,396,224
additional space required	5,477,329,803

The output status was BLOCKED_INSUFFICIENT_SPACE and its next action was:

python3 scripts/download_rtk_slam_dataset.py \
  --sequence construction_seq2 \
  --eval-assets \
  --dest /mnt/large/rtk_slam

The equivalent --dry-run --json returned a schema-stable plan with side_effects_started: false; both tested destination paths remained absent. No large DB3 was downloaded.

Attached-but-unmounted storage recovery

The next exact clean implementation carrier 0c3f58849b8ce7d8b4d464dbff891b244d781880 reproduced a second real user failure: a 2 TB SanDisk Extreme Portable SSD was attached, but its ext4 partition was not mounted. The former plan could only repeat a generic /mnt/large destination and left the operator to discover the device and mount path.

The new write-free plan detects sufficiently large hotplug filesystems through lsblk without mounting or probing their contents. On this host it reported:

Field	Exact observation
candidate device	`/dev/sda1`
model / filesystem	SanDisk Extreme Portable SSD / ext4
partition bytes	2,000,397,795,328
free-space state	`UNVERIFIED_UNTIL_MOUNTED`
required working bytes	11,886,726,027
root filesystem free bytes	6,339,293,184
root filesystem shortfall	5,547,432,843
next action	`udisksctl mount -b /dev/sda1`

Automatic mounting correctly did not occur. A direct mount attempt outside the helper stopped at the operating system's interactive authorization boundary; no password was supplied, no filesystem was mounted, and no download began. After the operator mounts the device, both follow-up commands are copy-ready and resolve the actual mount path themselves:

python3 scripts/download_rtk_slam_dataset.py \
  --sequence construction_seq2 \
  --eval-assets \
  --dest-device /dev/sda1 \
  --dry-run

# Run only after the command above reports READY.
python3 scripts/download_rtk_slam_dataset.py \
  --sequence construction_seq2 \
  --eval-assets \
  --dest-device /dev/sda1

The exact-carrier normal request against the undersized root filesystem again exited 2 before writes or network access and left its destination absent. A direct --dest-device /dev/sda1 --dry-run also exited 2 with the single mount action while the partition remained unmounted. No large DB3 was downloaded.

Small-asset and handoff validation

The pinned evaluation assets were fetched once into a generated temporary directory. The checkout was clean at the expected commit, contained the four required CSVs, and occupied 92,874,519 bytes including .git. This observation raised the explicit planning allowance from 50 MB to 150 MB; the separate minimum 1 GB reserve remains in force. A second dry-run selected reuse-verified with zero remaining payload.

The accuracy-suite preflight then accepted those checkpoint assets and rejected only the absent Construction Seq2 rosbag metadata. This proves the downloader and suite agree on their directory handoff without claiming a trajectory, accuracy result, or completed release profile. The 90 MB checkout and 36 MB MkDocs site used for validation were deleted afterward; they were generated temporary data and are not recoverable.

Enforced behavior

--dry-run and --list perform no network access or filesystem writes;
large attached-but-unmounted hotplug filesystems are reported read-only, never mounted or content-probed by the helper;
--dest-device resolves exactly one current mountpoint and removes manual mount-path substitution;
candidate partition size never substitutes for free-space verification; the mounted filesystem must pass a second dry-run before live acquisition;
--json is accepted only with --dry-run and keeps stdout machine-readable;
live capacity failure occurs before dependency checks, directory creation, downloads, or Git fetches;
existing regular partial files reduce the exact remaining transfer;
complete files are reused only after exact size and SHA-256 verification;
oversized, same-size-but-wrong, non-regular, and symlinked inputs fail closed;
completed downloads are checked by exact size and SHA-256, not a 1% warning;
eval assets are accepted only at the pinned clean commit; and
acquisition readiness never substitutes for exact-head benchmark evidence.

The original focused downloader and suite regressions pass 16 / 16. The attached-storage follow-up expands the direct downloader suite to 14 / 14; combined RTK/docs focused checks pass 42 / 42, release/bundle checks pass 27 / 27, and changed-file Jazzy ament_flake8 passes. The complete product gate initially exposed only the expected missing CTest registration and stale publication inventory; those review-control failures are repaired in the evidence synchronization commit. The exact attached-storage evidence carrier complete rerun passes:

graph_based_slam: 1,466 passed / 13 skipped / 11 known ImageIO warnings;
lidarslam: 1,040 passed; and
combined maintained Python gate: 2,506 passed / 13 skipped.

The synchronized publication plan is PLAN_VALID_LOCAL_ONLY with 314 paths, seven slices, and inventory SHA-256 78dc0990244e7b47e9f6b89d6ae71bebd9e5341d055c5a760d6416058b29119a.

The clean attached-storage evidence carrier 3a38154e77311bc11bcfdafc1267263b756e7dbe also passes the canonical fail-closed release-bundle rehearsal. The checker built and reverified the candidate twice before retaining one archive. Both builds were byte-identical and contained 264 manifest files; the resulting 11,948,788-byte archive has SHA-256 46b20d1eecaf6ab665c17816db42e937e9cb50c88a8dc4340beda4f2e670cf27. The archived RTK helper lists all four immutable sequence identities when run directly from its tar member. This is local candidate evidence, not a published release asset, and it must not be reused for a later commit.

A follow-up closes the extracted-bundle UX gap: the curated inventory now includes scripts/download_rtk_slam_dataset.py, and the release-bundle regression invokes --list from the archived bytes. The documentation keeps the larger measured suite bound to an exact source checkout and compatible built ROS workspace rather than presenting the audit bundle as a complete runtime tree.

Remaining release work

Run the copy-ready command on a filesystem with enough free space, then execute scripts/run_rtk_slam_accuracy_suite.py for Construction Seq2 and Construction Seq1 at the exact release candidate. Until both fresh outputs pass their profile contracts, all four current hard-profile blockers remain unchanged:

newer_college_math_hard;
ntu_viral_tnp_01;
mid360_gt_rtkslam_construction_seq2; and
mid360_gt_rtkslam_construction_seq1.

This work grants no tag, Release, image, merge, environment, benchmark claim, or dataset-publication authority.