refactor(algo): rename advantage component → algo.type

"advantage" was too general for a component that also fixes loss routing
and the model role, so flatten the `algo` bundle into a discriminated union
keyed on `type`: `[orchestrator.algo] type = "grpo"` instead of
`[orchestrator.algo.advantage] type = ...`.

- AlgorithmConfig is the union; each *AlgorithmConfig variant carries `type`,
  its params, and the shared sampling/teacher fields (BaseAlgorithmConfig).
- `algo.teacher` shorthand preserved (folds into `model` for opd/opsd,
  `sampling.source` for sft); `algo.model` is opd/opsd's direct field.
- Deleted the unused `advantage = {}` env/top-level shorthand + fold validator.
- Runtime: Algorithm takes the algo config (self.config); build_algorithm
  dispatches on algo.type. Migrated 12 TOMLs, docs, and unit tests.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: hallerite

configs/ci/integration/reverse_text_rl_opd/start.toml

@@ -18,7 +18,7 @@ name = "ci-rl-opd"
 batch_size = 128
 group_size = 16
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "opd"
 
 [orchestrator.algo.teacher]

configs/ci/integration/reverse_text_rl_sft/start.toml

@@ -18,7 +18,7 @@ name = "ci-rl-sft"
 batch_size = 128
 group_size = 16
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "sft"
 
 [orchestrator.algo.teacher]

configs/debug/algorithms/echo.toml

@@ -18,10 +18,10 @@ group_size = 4
 
 # alphabet-sort's feedback arrives as user messages, so train the user role
 # instead of echo's tool default.
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "echo"
 
-[orchestrator.algo.advantage.roles.user]
+[orchestrator.algo.roles.user]
 alpha = 0.1
 
 [[orchestrator.train.env]]

configs/debug/algorithms/grpo.toml

@@ -12,7 +12,7 @@ name = "debug-rl"
 batch_size = 128
 group_size = 16
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "grpo"
 
 [orchestrator.renderer]

configs/debug/algorithms/max_rl.toml

@@ -12,7 +12,7 @@ name = "debug-max-rl"
 batch_size = 128
 group_size = 16
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "max_rl"
 
 [orchestrator.renderer]

configs/debug/algorithms/mixed_grpo_opd.toml

@@ -22,7 +22,7 @@ name = "debug-mixed-grpo-opd"
 batch_size = 128
 group_size = 16
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "grpo"
 
 [orchestrator.renderer]
@@ -39,7 +39,7 @@ name = "reverse-text-grpo"
 id = "reverse-text"
 name = "reverse-text-opd"
 
-[orchestrator.train.env.algo.advantage]
+[orchestrator.train.env.algo]
 type = "opd"
 
 [orchestrator.train.env.algo.teacher]

configs/debug/algorithms/opd.toml

@@ -19,7 +19,7 @@ name = "debug-opd"
 batch_size = 128
 group_size = 16
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "opd"
 
 [orchestrator.algo.teacher]

configs/debug/algorithms/opd_lora.toml

@@ -19,7 +19,7 @@ name = "debug-opd-lora"
 batch_size = 128
 group_size = 16
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "opd"
 
 [orchestrator.algo.teacher]

configs/debug/algorithms/self_distill.toml

@@ -19,7 +19,7 @@ batch_size = 32
 group_size = 1
 
 # reverse-text's demo lives in the "answer" column.
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "opsd"
 model = "policy"
 demo_key = "answer"

configs/debug/algorithms/sft_distill.toml

@@ -19,7 +19,7 @@ name = "debug-sft"
 batch_size = 128
 group_size = 4
 
-[orchestrator.algo.advantage]
+[orchestrator.algo]
 type = "sft"
 
 [orchestrator.algo.teacher]