bot retraining. bots get confused on procedural maps. need to train them to b… #177

No description.

Planner, for whoever is holding this: the issue is one sentence, so here is what I would call done — take or leave any of it, it is a spec, not a rule.

• [ ] Say what "confused" is, in a number. Before changing anything, measure: seeded duels and seeded rounds on de_forge and on a handful of named generated seeds, win rate against the scripted baseline on each. That is #154's harness, and the measurement is worth landing even if the retraining is not.
• [ ] Train on the maps the game plays: a level drawn per episode — de_forge some of the time, a generated map the rest — from a seed stream the run records, so a run is reproducible from its own log. That is #156.
• [ ] The observation, the heads and OBS_SIZE do not change. Same network, more worlds. If the bots are confused because they cannot see what they need to (a wall reading, a site marker), that is a different and bigger issue — say so and open it rather than widening the observation here.
• [ ] The policy may not get worse on de_forge. If it is, the pull reports the numbers and the weights are not swapped in; a negative result plainly stated is a good outcome.
• [ ] The card in src/client/weights.ts carries both win rates, the page and the terminal show them, and weights.test.ts pins the new bytes.
• [ ] The run fits a session: the shipped weights cost about twelve minutes on one core for 8.2M ticks. State what this one cost.

#156 now waits on this issue rather than competing with it, and may well be closed by it. #154 stays open on its own: a measurement that outlives the retraining is worth having.