Document Koide acceptance reconfirm and Boreas baseline smoke.

Record run-to-run variance on Koide 120s/180s after recovery_retry tuning
and capture the Boreas 120s local_map_crop throughput bottleneck.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: rsasaki0109

CHANGELOG.md

@@ -19,7 +19,9 @@
 - document Istanbul public demo RMSE run variance; identical seed/map can yield ~`1.2 m` or late-run drift outliers
 - tune public Istanbul 60 s preset with `local_map_crop` and `recovery_retry_from_last_pose` (`r3`, `gap<=1 s`, `seed<=15 m`)
 - reconfirm public regression suite after drift tuning; Istanbul gate `1.00 m` and HDL IMU checks pass (2026-06-10)
-- enable `recovery_retry_from_last_pose` on Koide outdoor manifests; `180 s` acceptance rises from `71` to `473` ok rows
+- enable `recovery_retry_from_last_pose` on Koide outdoor manifests; `180 s` acceptance rises from `71` to `473+` ok rows
+- reconfirm Koide outdoor manifests after recovery tuning; acceptance holds on `180 s` reconfirm (`486/753 ok_rows`) with run-to-run RMSE variance
+- fix `boreas_localization_120s.yaml` dataset paths; baseline smoke shows `37/961 ok_rows` with `local_map_crop_too_small` bottleneck
 - guard `use_odom` integration until initial pose is valid and keep pose finite for NDT init guess
 - ignore non-finite `/initialpose` payloads and stop eager scan replay during pose reset
 - v1.1 relocalization is documented as a validated dry-run `/initialpose` command artifact only

docs/benchmarking.md

@@ -113,10 +113,12 @@ ros2 run lidar_localization_ros2 benchmark_from_manifest \
 
 Latest local Koide snapshot:
 
-- `2026-06-10`, `outdoor_hard_01a_smoke60` (baseline): `translation_rmse_m=0.230`, `rotation_rmse_deg=2.607`, `ok_rows=228/229`
-- `2026-06-10`, `outdoor_hard_01a_120` (baseline): `translation_rmse_m=0.192`, `rotation_rmse_deg=1.923`, `ok_rows=441/461`
+- `2026-06-10`, `outdoor_hard_01a_smoke60` + `recovery_retry r3_gap1_seed15`: `translation_rmse_m=0.228`, `rotation_rmse_deg=2.696`, `ok_rows=244/246`
+- `2026-06-10`, `outdoor_hard_01a_120` + `recovery_retry r3_gap1_seed15` (tuning run): `translation_rmse_m=0.192`, `rotation_rmse_deg=1.923`, `ok_rows=441/461`
+- `2026-06-10`, `outdoor_hard_01a_120` + `recovery_retry r3_gap1_seed15` (reconfirm run): `translation_rmse_m=0.315`, `ok_rows=221/447` — `fitness_exploded` outlier ~`104 s`
 - `2026-06-10`, `outdoor_hard_01a_180` (baseline): `translation_rmse_m=0.210`, `ok_rows=71/523` — pose output stops ~`126 s`
-- `2026-06-10`, `outdoor_hard_01a_180` + `recovery_retry r3_gap1_seed15`: `translation_rmse_m=0.255`, `rotation_rmse_deg=2.273`, `ok_rows=473/474`
+- `2026-06-10`, `outdoor_hard_01a_180` + `recovery_retry r3_gap1_seed15` (tuning run): `translation_rmse_m=0.255`, `rotation_rmse_deg=2.273`, `ok_rows=473/474`
+- `2026-06-10`, `outdoor_hard_01a_180` + `recovery_retry r3_gap1_seed15` (reconfirm run): `translation_rmse_m=1.549`, `rotation_rmse_deg=4.373`, `ok_rows=486/753` — acceptance holds, late-run drift outlier
 - `2026-06-10`, `outdoor_hard_01 window_15_120` recovery compare:
   - baseline: `translation_rmse_m=5.477`, `matched=138`
   - `recovery_retry r3_gap1_seed15`: `translation_rmse_m=0.266`, `matched=443`
@@ -333,6 +335,20 @@ Start from:
 The generated bundle contains mapping, reference extraction, localization benchmark, and Nav2 map
 generation scripts. See [mapless_public_dataset_workflow.md](mapless_public_dataset_workflow.md).
 
+### Boreas baseline snapshot
+
+Latest local Boreas snapshot (`boreas_localization_120s.yaml`, `boreas-2021-09-02-11-42` vs GT-aligned map):
+
+- `2026-06-10`, `120 s` localizer-only: `translation_rmse_m=43.515`, `rotation_rmse_deg=8.581`, `ok_rows=37/961`
+- dominant failure modes: `fitness_score_over_threshold_rejected` (`744`), `local_map_crop_too_small` (`179`), `accepted_gap_reinit_requested` from ~`43 s`
+- pose throughput: only `38` published poses in `120 s` — treat as map/crop/throughput bottleneck, not a passing benchmark
+
+Next engineering targets:
+
+1. verify map split / GT alignment for localization sequence vs mapping sequence
+2. disable or widen `local_map_crop` for large outdoor maps until crop policy is dataset-aware
+3. stage IMU/twist prediction only after localizer-only acceptance is stable
+
 ### Boreas starter
 
 Convert a Boreas raw sequence to rosbag2, then scaffold the split mapping/localization bundle.

docs/public_validation_log.md

@@ -3,6 +3,83 @@
 This file records reproducible public-data validation snapshots. It is not a
 hardware field-test log.
 
+## 2026-06-10 Boreas Baseline Smoke
+
+- commit: `4dd6b4c` (manifest path fix uncommitted)
+- source branch: `main`
+- validation command:
+
+```bash
+source scripts/setup_local_env.sh
+ros2 run lidar_localization_ros2 benchmark_from_manifest \
+  --manifest param/benchmark/boreas_localization_120s.yaml
+```
+
+Result:
+
+- `boreas-2021-09-02-11-42` `120 s` localizer-only: diagnostic failure
+  - translation RMSE: `43.515 m`
+  - rotation RMSE: `8.581 deg`
+  - matched samples: `37`
+  - ok rows: `37 / 961`
+  - dominant diagnostics: `local_map_crop_too_small`, `fitness_score_over_threshold_rejected`, `accepted_gap_reinit_requested`
+
+Artifacts:
+
+- `/tmp/lidarloc_boreas_localization_120s_throttle/`
+
+Interpretation:
+
+- Boreas is not ready for outward claims; current bottleneck is map/crop acceptance and very low pose throughput.
+- Next step is map-split validation and crop-policy tuning before IMU prediction work.
+
+## 2026-06-10 Main After Koide Outdoor Acceptance Tuning
+
+- commit: `4dd6b4c`
+- source branch: `main`
+- validation commands:
+
+```bash
+source scripts/setup_local_env.sh
+ros2 run lidar_localization_ros2 benchmark_from_manifest \
+  --manifest param/benchmark/koide_hard_localization_outdoor_hard_01a_smoke60.yaml
+ros2 run lidar_localization_ros2 benchmark_from_manifest \
+  --manifest param/benchmark/koide_hard_localization_outdoor_hard_01a_120.yaml
+ros2 run lidar_localization_ros2 benchmark_from_manifest \
+  --manifest param/benchmark/koide_hard_localization_outdoor_hard_01a_180.yaml
+```
+
+Result:
+
+- `outdoor_hard_01a_smoke60` + `recovery_retry r3_gap1_seed15`: `pass` acceptance
+  - translation RMSE: `0.228 m`
+  - rotation RMSE: `2.696 deg`
+  - matched samples: `244`
+  - ok rows: `244 / 246`
+- `outdoor_hard_01a_120` + `recovery_retry r3_gap1_seed15`: outlier on reconfirm
+  - translation RMSE: `0.315 m`
+  - matched samples: `221`
+  - ok rows: `221 / 447`
+  - failure mode: `fitness_exploded_reinit_requested` ~`104 s`, unrecovered
+- `outdoor_hard_01a_180` + `recovery_retry r3_gap1_seed15`: acceptance improved, accuracy outlier
+  - translation RMSE: `1.549 m` (tuning run was `0.255 m`)
+  - rotation RMSE: `4.373 deg`
+  - matched samples: `486`
+  - ok rows: `486 / 753` (baseline without recovery was `71 / 523`)
+
+Artifacts:
+
+- `/tmp/lidarloc_koide_outdoor_hard_01a_smoke60/`
+- `/tmp/lidarloc_koide_outdoor_hard_01a_120/`
+- `/tmp/lidarloc_koide_outdoor_hard_01a_180/`
+
+Interpretation:
+
+- `recovery_retry_from_last_pose` fixes the `180 s` acceptance collapse (`71 -> 473+ ok_rows`).
+- Koide outdoor replay shows run-to-run variance on `120 s` / `180 s` similar to Istanbul;
+  treat acceptance gains as the validated improvement and prefer repeat runs for release checks.
+- Full-route robust tracking on Koide `Outdoor01` remains an open engineering target.
+
 ## 2026-06-10 Main After Istanbul Drift Tuning
 
 - commit: `f6503d3`