critic family: refuse positive_lm_source=success without a success-writing reward function

[michaellchung]

Why

S19 #6 VAPO G3 s19-vapo-g3-20261010a failed with 0 rounds. Both islands raised at the first critic step: --positive-example-source success needs sample.metadata['success']. The G3 script loaded the out-of-tree gsm8k_reward:score, which writes no success flag. dry-run did not catch it.

S13 VAPO G1 (s13-forkg1-vapo-20261007e) passed with the same positive_lm_source=success, because it used the in-package yeto.rl.gsm8k_reward:score, which already writes metadata['success'] (since 4f1c03e).

Change

• New yeto/rl/reward_declarations.py: SUCCESS_WRITING_REWARDS lists the in-package Miles rewards that already write metadata['success'] (yeto.rl.gsm8k_reward:score, yeto.rl.math_reward:reward_func); writes_success decorator for user rewards. The reward modules themselves are unchanged (their source hashes are in the decoupling golden samples).
• New launch check positive_lm_success (critic.py): with positive_lm_coef set and positive_lm_source='success', rejects an importable reward function that is not declared. Sub-agent decision: an unimportable reference gets no verdict (test fixtures use placeholders; the Miles fork still refuses on the island). Skipped when the caller passes no reward function.
• reward_function is passed to the check from the launcher (dry-run / pre-launch) and the Miles island config.

Main agent decision (1+3): keep the VAPO spec semantics; fix the reward side and add the check. The G3 script should use --reward-function yeto.rl.gsm8k_reward:score for CASE=vapo (script lives outside the repo).

Tests

• tests/test_rl_positive_lm_success_check.py: 6 passed (refuse undeclared / unimportable; in-package rewards pass; skip cases; gsm8k success equals reward == 1.0).
• G3 dry-run with this commit: gsm8k_reward:score -> rc=1 [positive_lm_success] rejection; yeto.rl.gsm8k_reward:score -> rc=0.
• Local safe suite (-m "not gpu and not ray_local"): 5384 passed, 47 skipped, 1 failed: test_codex_bridge_compaction.py::test_off_stock_bridge_unchanged_and_compactionrl_without_trigger_is_byte_identical; that file passes alone on this branch and on main (21 passed), so it is order-dependent, not caused by this PR.
• 4cd3109 (G3 code) already contains 4f1c03e; the G3 failure came from the script's reward choice, not missing code.

No GPU run. No image rebuild needed: the yeto package ships with the workdir; the Miles fork in the image is unchanged.

🤖 Generated with Claude Code