Add synthetic data replay ratio support

[Oaklight]

Summary

• Add synthetic_ratio and synthetic_sources parameters to SamplerConfig
• Controls proportion of synthetic vs benchmark data per epoch (e.g. 0.4 = 40% synthetic, 60% benchmark)
• Inspired by CLM paper showing 40% replay prevents catastrophic forgetting
• Per-type allocation: splits each type's budget between synthetic and benchmark sources
• Graceful fallback: if a type has no synthetic sources, all items come from benchmark (and vice versa)
• Auto-detection: sources with "synthetic" in name/path are auto-tagged if synthetic_sources not explicitly set
• Wired through YAML config and CLI (--synthetic-ratio 0.4)
• 5 new tests covering ratio control, edge cases, and fallback behavior

Closes part of #97 (synthetic replay ratio approach)

Test plan

• python -m pytest tests/test_sampler.py -v — 14 tests pass (9 existing + 5 new)

[clementine-oaklight]

LGTM ✅ — Clean synthetic replay ratio support.

Component	Status
Split sampling	✅ _split_sample divides budget per type
Fallback	✅ Falls back to available group when one side empty
Auto-detect	✅ Sources with "synthetic" in name/path auto-tagged
CLI	✅ --synthetic-ratio + YAML config
Tests	✅ 5 tests (ratio control, edge cases, fallback)

CLM-inspired 40% replay approach. The temporary _source_weights swap in _split_sample is pragmatic — works correctly and avoids duplicating the weighted sampling logic.

CI green.

[milo-oaklight]

Approved ✅

Synthetic data replay ratio support (CLM-inspired):

Design

• synthetic_ratio: target proportion of synthetic data per type (0.0–1.0)
• synthetic_sources: explicit set of source names, or auto-detected
• Per-type split: each type's budget divided between synthetic and benchmark

Implementation

• _split_sample(): separates sources into syn/bench groups, draws proportionally
• Graceful fallback: if type has no synthetic sources → all from benchmark (and vice versa)
• Auto-detection: sources with "synthetic" in name/path tagged automatically

Wiring

• CLI: --synthetic-ratio 0.4
• YAML: sampling.synthetic_ratio, sampling.synthetic_sources

Tests (5 new)

• test_synthetic_ratio_controls_proportion: ~40% at ratio=0.4
• test_synthetic_ratio_none_no_effect: disabled by default
• test_synthetic_ratio_zero_excludes_synthetic: 0 synthetic items
• test_synthetic_ratio_one_all_synthetic: 100% synthetic
• test_synthetic_ratio_with_missing_synthetic_type: fallback per-type

CI green.

[elena-oaklight]

Summary

Clean implementation of synthetic data replay ratio for CLM-style data mixing. The PR adds a configurable synthetic_ratio parameter to control the proportion of synthetic vs. benchmark data per question type during sampling.

Reviewed Items

Component	Status	Notes
SamplerConfig fields	✅	synthetic_sources: set[str], `synthetic_ratio: float
_split_sample method	✅	Splits by source category, reuses _weighted_sample
sample_epoch integration	✅	Routes to split/weighted based on config
build_sampler_config	✅	Parses config + auto-detects synthetic sources
CLI --synthetic-ratio	✅	Float arg, overrides config
Config YAMLs	✅	Commented examples in all 5 configs
Tests	✅	5 new tests covering core logic + edge cases

Implementation Highlights

1. Backward compatible: synthetic_ratio=None (default) preserves existing behavior
2. Graceful fallbacks: If a type has no synthetic (or benchmark) sources, draws all from the available side
3. Integer rounding: n_synthetic = int(n * ratio), n_benchmark = n - n_synthetic ensures exact total
4. Auto-detection: Sources with "synthetic" in name or path are auto-classified

Minor Observation

The _split_sample temporarily mutates self._source_weights during sampling, then restores it. This works correctly because sample_epoch is single-threaded, but a comment noting this invariant would be helpful for future readers.

Not blocking — the pattern is safe as-is.

LGTM ✅