Strategy search 2026-10-11: five split_shape variants, LeadAndSupport shot fix

[isaac0804]

One session (2026-10-11/12): an orchestrated strategy search, plus the shared fixes it turned up.

Shared fixes (each with a regression test that fails before, passes after)

• LeadAndSupportTactic shot-target flip (d7659277): with two near-tied gaps the leader's target flipped every tick, so it never kicked and the match froze. Now uses _pass_and_score's 0.15 m switch margin.
• turn_on_spot blocked pivot (205ccfeb): when an enemy blocked the pivot, the final override replaced move()'s escape command with the pivot term and pushed into the blocker. A blocked pivot now keeps the escape command.
• DecoyOverloadTactic finish phase (d3b093b2): never released when the ball stopped making progress. Now releases after 3 s without progress, like its timeout.
• docs/STRATEGY_DEVELOPMENT.md: leftover merge-conflict markers removed (46e9fa4a).

These touch shared tactics and a shared skill, so check_strategy_branch.py --base main flags them by design.

Strategy search

Haiku subagents researched split_shape and implemented one-idea variants in their own worktrees. Each was reviewed and passed its own tests and test_all_strategy_configs.py. Six kept: ee647a3a, plus split_outlet_wall (bd71d436). split_wall's free-kick wall committed keep-out fouls when the box clamp pulled it within 0.5 m of the ball. Fixed in 47af142b: keep-out fouls fell from 115 to 14 for split_wall and from 58 to 7 for split_outlet_wall.

Round-robin, 18 configs, 153 matches, with all fixes (replays/tournament_20261012_095359):

Strategy	Idea	W-D-L	GF-GA	Points per match	vs split_shape
split_sticky	possession hysteresis	15-0-2	136-83	2.65	8-7
split_block	BlockShapeTactic defense	14-1-2	144-82	2.53	8-8
split_wall	two-robot wall at the opponent's direct free kicks (first strategy to use referee_overrides)	14-1-2	119-75	2.53	7-6
split_outlet	one robot held high on the weak side while defending (new OutletTactic)	14-0-3	127-85	2.47	8-7
split_outlet_wall	outlet plus wall	13-1-3	127-74	2.35	7-2
split_press	2-robot press without the ball	12-2-3	138-73	2.24	8-8
split_shape	reference	10-3-4	113-76	1.94

Every variant finished above split_shape, but the order among them changed with every shared fix. The screen is one deterministic match per pairing, so it can't rank them.

Head to head with seeded noise (2026-10-12). Each pairing played its 8 deterministic side/kickoff settings and 20 seeded noisy matches: side fixed, kickoff alternating, 3 mm and 1° vision noise, 2% of detections dropped. The noise comes from the existing rsim_noise/rsim_vanishing, injected by a scratch script, so no repo or match-cache change. Goal difference per match for the first-named strategy:

Pairing	8 settings, no noise	20 noisy seeds
split_block vs split_shape	6-2-0, +2.4 ± 0.9	11-1-8, +0.5 ± 0.6
split_outlet vs split_shape	4-0-4, −1.3 ± 1.3	8-4-8, −0.6 ± 0.8
split_sticky vs split_block	3-1-4, +0.0 ± 1.4	8-4-8, −0.3 ± 0.5
split_shape vs counter_flow	7-1-0, +4.0 ± 0.7	17-1-2, +3.1 ± 0.5

No variant is shown better than split_shape, which stays the champion. split_block was picked on its deterministic result, and that edge did not hold over 20 noisy matches. Noise keeps the real gap against counter_flow, a seed replays bit for bit, and each match has about the same spread as a deterministic one. So noise gives as many matches per pairing as needed, without making any single match noisier. That's the basis for the evaluation change (roadmap 2a), in a separate session.

The scenario bench (860 starts, 14 min) found no difference for split_block vs split_shape (t = +0.8). 63% of its starts tied.

Known open

6 of 153 round-robin matches stalled; none of the 80 noisy matches did. In each, two carriers pinned the ball between them until the referee's no-progress stop ended it. The cause of the stuck turn is not found (diagnosis in the session notes); fixing it changes a shared skill, so it is left for its own session.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ECXiwf3VYxBTMmrLaexRQK