M52: Crazy Fruits combo curriculum — seeded-specials starts + combo-biased exploration

[PieterjanDeClippel]

Why

Owner question after M51 shipped: has the net ever experienced what a wrapped+wrapped or bomb+bomb combo pays? Investigation answer: the deterministic combo payoff is already in the net's input planes and dense targets — what's missing is state frequency (combo-legal states ≈2.8% of natural play; adjacent bomb+bomb ≈ never) and the realized post-combo refill experience only playing them provides.

Full design + 2-agent verification record: docs/prd/CRAZY_FRUITS_COMBO_CURRICULUM_PRD.md (PLAN M52).

What

• Combo-biased ε-exploration (primary lever): generic DqnOptions.ExploreBias hook, consulted only inside the ε-branch (null ⇒ zero extra RNG draws ⇒ every other game bitwise-identical). Crazy Fruits wires it so an exploration step with a legal special+special swap takes one with probability q — realized combo experience at zero distribution shift.
• Seeded-specials training starts: CrazyFruitsEnv.SeedSpecialsProb — with probability p the training board is dealt combo-ready (one adjacent special pair + up to 2 singles) by overwriting plain cells' kinds. Fruit types are kept, so the deal's no-instant-match / has-legal-swap guarantees hold by construction; draws come from the env RNG stream (deterministic); default OFF — the eval env and every gate stay on natural boards. Exists specifically to manufacture adjacent bomb+bomb, which ε-bias can't reach.
• Adjacent pairs only — a near pair's bring-together swap has a flat γ=0 label (the multi-move setup skill stays with the future search-distillation lever).
• Lab: --seed-specials / --combo-explore; per-kind created counts in --baselines (new gate instrument).
• Tests: 67/67 green (4 new: seeded invariants, per-seed determinism, suggestion-hook legality, agent bias redirect).

Training status — 🔄 IN PROGRESS (draft until gates run)

cf10train = full resume of cf9train's training state (Adam moments + 100k natural replay buffer carry over; owner choice: reuse checkpoint + full 400k budget) → 800k absolute steps, p=0.25, q=0.5. Keep-best is armed at the resumed net's own eval, so the deployable net is only overwritten when beaten.

Pre-registered gates (run when training completes):

1. 500-ep natural-board score NOT CI-separated below cf9train's 5666.4
2. Probe (1000 eps, natural boards): combo take-rate ≥ 75% (baseline 63%)
3. Bomb-bucket opportune take ≥ 85% (baseline 80%)
4. All M50.3 bars stay green (≥+30%/random · gap-share ≥64% · created ≥7.3 / fired ≥5.6)
5. Per-kind created non-regression vs cf9train, especially bomb-created

Escalation if 2/3 miss: raise q, one run, stop-loss. This PR ships the new checkpoint only if gates pass; the code is complete and independently tested either way.

🤖 Generated with Claude Code