[rsasaki0109]

Summary

• What changed?
  • gnss_fgo_imu_no_base now accepts --native-course-attitude-seeds-mount-deg together with --native-source-imu-initialization.
    • The initial IMU pass takes the course seeds instead of the source heading seeds.
    • The refinement pass turns course seeds off and takes the initial pass's optimized attitude, which it hands over as heading seeds.
    • Every other flag combination builds the same configuration as before.
  • The GSDC official recipe now rebuilds Kaggle submission 57678818 (Private 0.806 m / Public 0.816 m). The previous one was 57664241 (0.811 / 0.817 m). The third pass is now on 32 drives; these two are new:
    • sjc-r A505U, which keeps its course seeds;
    • lax-hh A325G.
  • The recipe's binary moves to the new build. The whole lane was rerun on it from scratch, byte-identical (see Validation).
• Why?
  • sjc-r A505U, a Private drive, is mounted backwards and needs its course seeds (178.9 deg). The third pass needs source IMU initialization, and the solver rejected that combination.

Scope

• Single feature / single operational improvement / single user-visible value
• Adds or tightens one regression, sign-off, dogfood, or artifact-schema check
• No unrelated refactor mixed in
• No docs-only noise mixed into solver/protocol work. The docs record the score change.
• Protocol ingest, parser work, runtime knobs, solver behavior, and docs are split unless this PR explicitly needs the boundary crossing. The solver change is one admission plus its config split, in its own commit.

External Reference

• Source repo used for comparison: taroz/gsdc2023 (fgo_gnss_imu.m). Its final pass starts from the previous pass's attitude unless RPYReset is set.
• What was adopted here: the same handoff for the course-seeded drive.
• What was intentionally not adopted:
  • mtv-pe1 A325G: Kaggle Public was worse with the pass, 0.821 m vs 0.817 m.
  • mtv-e Pixel4 XL: no difference.
  • Mi8: still worse on train.

Behavior Boundary

• Runtime default changed? No. The change only affects runs that pass both flags, which was a CLI rejection before.
• Env-gated or config-gated? Yes, by CLI flags.
• Artifact/schema changed? No schema change. The recipe's binary pin and two drive entries change.
• Dataset or credentials required? The lane needs the taroz dataset_2023. The CLI test is admission-only and uses no data.

Validation

• python3 tests/test_cli_tools.py
• python3 tests/test_benchmark_scripts.py
• python3 tests/test_packaging.py
• python3 tests/test_python_bindings.py -v
• python3 tests/test_ros2_node.py
• ctest --test-dir build --output-on-failure

Focused checks run:

pytest tests/test_smartphone_course_attitude_seeds_cli.py   # 7 passed on the new build; the admission test fails on the old one
python3 tests/test_gnss_reproduce.py                          # OK
# A505U recipe argv without the pass on the new build: byte-identical to its pinned output (d68137e1)
# full lane from scratch on the new build (cccb04cc), four at a time on 4 cores, 2 h 21 min:
#   25/25 stage-0 identical, height-maps 25/25 match, 40/40 final identical (lax-p single_cpu),
#   assemble -> 2e0deb2c... MATCH

Sign-off / Benchmark Impact

• Added or updated regression/sign-off:

  • The CLI test now admits course seeds with source IMU initialization and still rejects them with epoch heading seeds.
  • The tracked-recipe test pins 57678818 / 2e0deb2c and checks that 32 drives carry the refinement pass.

• Affected dataset(s): GSDC 2023-2024 test (official).

• Expected metric impact (Kaggle readback):

  Ref	Change	Public	Private
  57664241	before	0.817	0.811
  57677499	plus sjc-r A505U	0.817	0.806
  57677502	probe: plus mtv-e Pixel4 XL	0.817	0.806
  57677505	probe: plus mtv-pe1 A325G	0.821	0.806
  57678818	plus lax-hh A325G (recipe)	0.816	0.806

• Artifact(s) produced: none tracked.

Notes

• Risk: low. The new code path is reachable only with a flag pair that used to be rejected.
• Follow-up PRs: the Mi8 drives.
• Explicit non-goals: changing heading-seed behavior on any other drive.

---

Generated by Claude Code