Update development plan with Jazzy-green public suite and execution log

Records the first Jazzy-green public regression numbers (Istanbul 0.673 m
/ 0.036 s median, HDL 468.5/459.0 rows), the shared-machine throughput-gate
lesson, the --resume failed-output trap, the Nav2 occupancy-map bootstrap
prerequisite on fresh workspaces, and the BBS_2D landing criteria for the
global localization G1 track.

Author: rsasaki0109

docs/development_plan.md

@@ -25,7 +25,13 @@ What is true right now:
   map-alignment/covariance documentation set, Istanbul drift tuning, Koide outdoor recovery
   tuning, and Boreas diagnostic work.
 - The 2026-05-22 release regression snapshot (`overall_pass=true`) plus the 2026-06-10
-  drift-tuning reconfirm are the latest recorded green boundaries.
+  drift-tuning reconfirm are the latest recorded green boundaries on the Humble overlay.
+- New on 2026-06-11: the public regression suite is green on a ROS 2 Jazzy build for the
+  first time (`overall_pass=true`): Istanbul 60 s no-IMU translation RMSE `0.673 m`,
+  rotation RMSE `2.179 deg`, alignment median `0.036 s`; HDL 60 s two-repeat medians
+  `468.5` baseline / `459.0` IMU-candidate pose rows, IMU-candidate alignment median
+  `0.084 s`, fallback events `1`. The long Nav2 reinitialization-supervisor regression is
+  the remaining release-suite stage (see execution log below).
 - The P0 crash issues are fixed on `main` but still open on GitHub, which makes the
   package look more broken than it is to a new user. Closing them is gated on tagging the
   release that contains the fixes.
@@ -45,6 +51,13 @@ Known execution constraints:
 - Jazzy builds still use the `ndt_omp_ros2` `humble` branch, matching CI.
 - `run_release_regression_suite.sh --resume` reuses failed run outputs; delete the failed
   subdirectories from the work dir before resuming.
+- The Nav2 reinitialization-supervisor regression depends on a workspace-local artifact
+  that is not shipped in the repo: `artifacts/public/autoware_istanbul_60s_nav2_map/
+  istanbul_60s_nav2_map.yaml`. On a fresh workspace it must be bootstrapped once from the
+  Istanbul map with `scripts/generate_occupancy_map_from_pcd.py --pcd
+  data/official/autoware_istanbul/pointcloud_map.pcd --output-dir
+  artifacts/public/autoware_istanbul_60s_nav2_map --map-name istanbul_60s_nav2_map`.
+  Folding this bootstrap into the regression entry script is an open cleanup item.
 
 ## Phase 0: v1.1.0 Release Closeout
 
@@ -54,9 +67,11 @@ the public state of the project matches the actual state of `main`.
 Steps:
 
 1. Run `scripts/run_release_regression_suite.sh` on an idle machine and confirm
-   `overall_pass=true`. Status: blocked on machine availability; build, unit tests, and
-   experiment suite already pass; Istanbul RMSE passed even under load (1.39 m vs 6.0 m
-   gate) while throughput gates did not.
+   `overall_pass=true`. Status: the public-suite half is green on Jazzy as of 2026-06-11
+   (numbers in Current State); the Nav2 reinitialization-supervisor half is pending the
+   occupancy-map bootstrap described in the execution constraints. Earlier same-day
+   attempts under load ~37 failed only on throughput gates, not correctness; Istanbul RMSE
+   passed even then (1.39 m vs the 6.0 m gate).
 2. Move the `Unreleased` section of `CHANGELOG.md` to `1.1.0`, bump `package.xml` to
    `1.1.0`, record the validation snapshot in
    [public_validation_log.md](public_validation_log.md), and tag `v1.1.0`.
@@ -141,9 +156,12 @@ Runs in parallel with Phases 1-3; detailed phases live in
 
 - G1 (artifact-first map-wide candidates): `MAP_GRID` baseline implemented 2026-06-11
   (unit-tested, verified on the Istanbul map and against the registration job generator).
-  Next: evaluate it on the Koide 180 s window against the route-proximity artifact, then
-  implement `BBS_2D` to cut the candidate count by orders of magnitude, then `FPFH_RANSAC`
-  as the full-3D fallback.
+  A `BBS_2D` implementation (multi-resolution occupancy pyramid with admissible
+  branch-and-bound over x/y/yaw, scan input from the bag at the request-window trigger) is
+  being drafted and reviewed; it lands only with passing unit tests for bound
+  admissibility and top-K ordering. Next after that: evaluate both engines on the Koide
+  180 s window against the route-proximity artifact, then `FPFH_RANSAC` as the full-3D
+  fallback.
 - G2 (runtime query service): opt-in service returning ranked candidates; starts after G1
   beats or matches route-proximity on the validated windows.
 - G3 (guarded automatic reinitialization): connects `/reinitialization_requested` to G2
@@ -164,6 +182,51 @@ characterization, Jetson + MID-360 hardware validation as a separate track.
 - BSD/MIT/Apache sources only for direct adaptation (e.g. `hdl_global_localization` is
   BSD-2 and safe to reference).
 
+## Execution Log
+
+### 2026-06-11: Jazzy bring-up and first Jazzy-green public suite
+
+What happened, in order, on a fresh Jazzy-only workspace:
+
+1. Workspace bootstrap: `repo` + `build_ws` (with `ndt_omp_ros2` `humble` branch) +
+   `local_prefix`, per [local_build.md](local_build.md). `colcon build` succeeded; all 32
+   unit tests and the experiment suite passed.
+2. Two environment fixes were needed and are now on `main`:
+   - `setup_local_env.sh` hardcoded Humble; it now resolves
+     `LIDAR_LOCALIZATION_ROS_DISTRO` / `ROS_DISTRO` / humble / jazzy in that order.
+   - The HDL sample fetch patched rosbag2 metadata into the Humble-only string form,
+     which Jazzy's rosbag2 (metadata version 9) rejects with a yaml-cpp `bad conversion`
+     at `offered_qos_profiles`; the patch is now distro-aware.
+3. First release-regression attempt ran while the machine carried unrelated workloads
+   (load ~37 on 20 hardware threads). Every timing-gated check failed (HDL alignment
+   median `0.26-0.32 s` vs the `0.10 s` gate, pose rows ~95 vs 250) while correctness
+   held (processes survived, Istanbul RMSE `1.39 m` vs the `6.0 m` gate). Lesson recorded
+   in the execution constraints: performance gates are meaningless on a loaded machine.
+4. A `--resume` retry silently reused the broken HDL outputs and failed again in
+   seconds; the failed subdirectories must be deleted first. Also recorded above.
+5. After load dropped to ~7, a clean rerun produced the first Jazzy-green public suite:
+   `overall_pass=true`, Istanbul `0.673 m` / `2.179 deg` / alignment median `0.036 s`,
+   HDL medians `468.5 / 459.0` pose rows with IMU-candidate alignment median `0.084 s`.
+   These numbers are comparable to the recorded Humble boundary (`1.176 m` Istanbul,
+   `~550` HDL rows) and better on Istanbul RMSE for this run.
+6. The release suite then stopped at the Nav2 reinitialization-supervisor stage: the
+   occupancy-map artifact `istanbul_60s_nav2_map.yaml` is workspace-local and absent on a
+   fresh checkout. Regeneration from the public Istanbul PCD is in progress; folding the
+   bootstrap into the entry script is an open cleanup item.
+
+Release decision: `v1.1.0` tags only after the full release suite (public + Nav2
+supervisor) is green in one run. The public-suite half is no longer the blocker.
+
+### 2026-06-11: Global localization G1 start
+
+- `MAP_GRID` baseline merged: map-wide seed candidates (occupied-cell centroids with a
+  per-cell ground-percentile z and discretized yaw) feeding the unchanged v1.1
+  registration-scoring and dry-run command chain. Verified on the Istanbul map (21,172
+  cells at 20 m spacing, deterministic striding to the candidate cap) and against
+  `make_registration_relocalization_jobs.py` with `candidate_index` ordering.
+- `BBS_2D` drafting started the same day (see the track section above for the landing
+  criteria).
+
 ## Suggested Order Of Work
 
 1. Phase 0 (release + issue hygiene) — small, high leverage, mostly waiting on an idle