Patch v2.2.3 - Kick Pipeline Fixes + RL Kick Hand-off + GlanceAtGoal

v2.2.3 - Kick Pipeline Fixes + RL Kick Hand-off + GlanceAtGoal

Fixes the two biggest on-field problems observed with v2.2.2: the robot hesitating ~15 seconds before kicking, and kicks going visibly off target even while the camera tracked the ball.

What changed

Kick hesitation (P0)

• Fixed the inverted goalpost-margin fallback in Brain::isAngleGood() — the strict 0.5m margin applied whenever the goal window was smaller than 120° (i.e. almost always), leaving a ±6-8° shot-acceptance window at range. The big margin now only applies close to the goal.
• Relaxed reachedKickDir (static threshold 0.1 → 0.2 rad) while tightening the sign-flip release (π/6 → 0.25 rad).
• Added a 2.5s kick-decision commit latch in StrikerDecide so the robot's own motion can no longer flap the angle conditions and halt the stateful Kick node mid-execution.
• Fixed the "stabilize" phase of Kick: applyMinX amplified a -0.05 m/s hold command into a 0.3 m/s backward walk that pushed the ball out of kick range. Now stands still.
• Retunes: near_ball_speed_limit 0.4 → 0.6, Adjust tangential_speed_near 0.15 → 0.3, kick gate tolerates ≤400ms ball-detection dropout at contact range.

Kick direction (P0)

• Kick now crab-walks through the ball along the planned kickDir (aiming 0.3m beyond the ball, clamped ±0.35 rad of the ball bearing) instead of walking at the raw ball bearing.
• GoalieDecide writes its clear direction into data->kickDir so goalie kicks aim correctly.

RL VisualKick band split — enable_auto_visual_kick re-enabled

• The RL branch now only triggers in a 1.2-4.0m band (code-enforced floor of kick_range * 1.2), so it can never shadow the deterministic Kick window again. RL owns the approach; the scripted kick owns the finish.
• Stall hand-off: if the ball sits within 0.85m essentially unmoved for 2.5s ("track-but-never-kick" firmware behavior), RLVisionKick exits and the scripted kick takes over.
• Implemented the previously stubbed isMinIntervalSatisfied() as a real 4s re-entry cooldown — no more exit/re-enter thrash.

New: GlanceAtGoal

• Before a kick/safe_shoot, if no opponent-half goalpost was seen and no successful re-localization happened within 4s, the robot pauses ~0.6s and looks at the goal to acquire goalposts / re-localize, then shoots. Config: strategy.glance_at_goal.*.

Other fixes

• Removed the empty-decision goalie freeze branch (ball 0.5-2.0m ahead caused no-op ticks).
• Possession detection now measures opponent distance from vision-confirmed Opponent robots instead of raw depth-grid obstacles (teammates/referees no longer hijack the kick into steal/defend).
• Steal-vs-defend risk roll cached 1s instead of re-rolled every 10ms tick; lead election got a sticky margin; onHalted seconds-vs-ms timer bugs fixed; duplicated leg-selection block removed.

Packaging

• Repository source folder v2.2.3/ excludes TensorRT .engine model files.
• Release asset v2.2.3.zip contains the full patch folder, including model engines.

Validation

• All changes cross-checked for symbol/registration/config consistency (new BT node registered, params declared, XML wired).
• Full ROS/colcon build was not run on this PC because ROS/colcon is not installed locally — build with ./scripts/build_brain.sh on the robot before deploying.
• Field-test tip: watch the debug/RLVisionKick rerun channel — exit reasons show whether the RL policy actually kicks in the band (ball_too_far exits) or relies on the stall hand-off (stall_handoff exits).