feat(orchestrator): task sampler skeleton

[hallerite]

The TaskSampler foundation in one PR, three commits: free the name, build the memory, retire the filter pipeline. Picks are byte-for-byte identical to main by default — this is the observability half of the curriculum design (design/task-sampler.md): estimates get warm and visible before anything is allowed to steer sampling.

1. chore(orchestrator): rename Sampler to RolloutSource — pure rename. The class is one env's rollout source (which model generates its train rollouts); its docstring already said so. Frees "sampler" for the task-sampling layer.

2. feat(orchestrator): TaskSampler skeleton with per-task outcome stats — TrainSource becomes TaskSampler: same pick contract to the dispatcher, plus a memory.

• task_stats.py: discounted success/failure counts (Beta posterior with forgetting) keyed by (env, content-hash task key, agent role), plus reward mean/std EMAs. Keys recompute from the trace echo, so stats survive dataset reordering and resumes. Nothing is evicted — estimates decay back toward the prior.
• TrainSink.on_group_finalized: one callback per finalized train group (all-errored ones included). Sink stays sampler-blind; sampler stays token-blind.
• sampler/{env}/* metrics: pool occupancy (learnable/saturated/hopeless/unseen), coverage, realized_signal_rate, wasted_token_frac (inference tokens in zero-signal groups — the number later enforcement has to beat).
• Checkpointing: stats ride inside the existing "train_source" key; older checkpoints load with empty stats and warm back up.

3. feat(orchestrator): retire the filter pipeline — the filter subsystem served three detectors; each gets a proper home and the pipeline (both config slots, double apply_filters pass, RolloutFilter protocol) is deleted. Group finalization becomes the pipeline's single decision point.

• Zero-advantage splits into its three roles: evidence (groups feed the sampler), admission ([orchestrator.sampler] drop_degenerate_groups = true drops zero-signal groups so the batch backfills — the old pre-batch enforce + oversampling recipe), and the ship rule (Algorithm.trains_on_zero_advantage, default False, echo True — MaxRL's all-zero sentinel stops depending on a filter being configured; advantages=None always ships).
• Gibberish/repetition demote to rollout-granularity detections ([[orchestrator.detections]]), evaluated at group finalization. An enforced detection ships no samples and never occupies a batch slot; its reward still counts toward the group baseline.
• Rollout.filter_results/is_filtered → detections/is_excluded; pre_filters/* metrics → detections/*; five tracked configs migrated, no shim. Defaults map 1:1. One deliberate metrics change: zero-advantage rollouts stay in the effective subset (excluding all-correct groups clipped reward stats).

Overlaps in intent with #3254 (independently converged on trains_on_zero_advantage, killing the filter slots, and drop-with-backfill); diverges on configurability, drop granularity, and the evidence wire — reconciliation discussed there.

Verified: full non-GPU unit suite passes at each commit (143 orchestrator+config tests at the tip, incl. 5 new task_stats tests and the rewritten test_detections.py); known-environmental failures (qwen3_vl_e2e, test_load_configs, reverse_text) unaffected.

🤖 Generated with Claude Code