Fixes #2445.

Context compaction can turn one sampled rollout into multiple training rows. BatchBuilder._post_process_rewards currently normalizes the whole batch when the row count differs from rollout_batch_size * n_samples_per_prompt. This mixes unrelated queries. Grouping rows by query alone also overweights episodes that produce more rows.

This change groups by Sample.group_index, computes the mean and optional sample standard deviation over distinct rollout_id values within each query, then broadcasts each episode's normalized reward to its rows in their original order. For example, query A with episode rewards [1, 0] and query B with [0, 0] keep centered rewards [0.5, -0.5] and [0, 0], even when episodes are split during compaction.

• Preserve fixed-size positional grouping for legacy generators that omit all query IDs; reject partially missing IDs or detectable ambiguous fan-out.
• Treat missing rollout IDs as independent rows, consistent with BatchBuilder.convert.
• Return zero for singleton groups and reject conflicting rewards within an episode. Custom reward postprocessing remains available for segment-specific reward semantics.
• Cover the existing GRPO, GSPO, CISPO and REINFORCE++ baseline normalization modes without changing loss reduction or token masks.

Validation: 86 CPU tests passed, including 27 new reward regressions and the existing trajectory-manager branching, Sample, rollout-data and sample-hook tests. The regression exercises the real TrajectoryManager -> BatchBuilder.convert -> get_grpo_returns path with and without compaction. On the unmodified base, 19 of the 27 new cases fail. The new test file is registered in CPU CI; Black 24.3.0, isort 5.13.2, Ruff 0.14.7 and git diff --check pass. No GPU training run was performed.

Related work: #2062 already proposes the same query/rollout grouping principle; #2204 addresses explicit query grouping. This is a focused implementation against the current BatchBuilder, with native compaction regression coverage and strict handling of ambiguous legacy batches. Both earlier PRs remain unmerged as of this submission preparation.