[doctorMcy]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

This PR adds unit tests for src/twinkle/metric/: 77 new test functions (91 → 168) across six files under tests/metric/:

• test_grpo_clip_sampler.py (new, 27): GRPO / CISPO / GSPO clip decisions, sampler reconciliation, top-k high-KL records, temperature rescaling, multi-microbatch cursor and input edge cases.
• test_metrics.py (66 → 96): Accuracy seq strategy, LossMetric sum reduction and per-channel aggregation, TrainMetric, DPOMetric ragged ref_logps, EmbeddingMetric, GRPO old-logps row mismatch, reporting helpers.
• test_ppo_metric.py (16 → 18): PPOValueMetric explained variance / clip / advantage stats, and PPOMetric parity with GRPOMetric.
• test_reporting.py (4 → 15): JSONL append and flush, timeouts and non-fatal backend failures, SwanLab prefix.
• test_throughput.py (5 → 9): variable-length batches and padded width, invalid input rejection, window throughput and memory reduction.
• test_types.py (new, 3): MetricRecord immutability and stage/status validation.

Two supporting source fixes:

• GRPOMetric did not advance its cursors when a microbatch was dropped. The dropped microbatch still owns rows in the flattened old / adv / sampler lists, so later microbatches were aligned against the wrong old logps and the global sample index in high-KL records shifted; empty logps also raised TypeError via len(None). Both skip cases now advance cursor / _gsi_cursor by the microbatch's logps row count before continue, and empty logps returns early.
• TrainMetric now guards the reported speed against a zero-length timing window. Coarse platform clocks (Windows) can return the same timestamp twice, and step delta / 0 raised ZeroDivisionError inside metric calculation; a zero window now reports 0.0, leaving iters and elapsed time unchanged.

PR information

本 PR 为 src/twinkle/metric/ 补充单元测试，新增 77 个测试函数（91 → 168），分布在 tests/metric/ 下 6 个文件：

• test_grpo_clip_sampler.py（新增 27 个）：GRPO / CISPO / GSPO 的 clip 判定、与采样端对账、top-k 高 KL 记录、温度重标定、多 microbatch 游标与输入边界处理。
• test_metrics.py（66 → 96）：Accuracy 的 seq 策略、LossMetric 的 sum 归约与分通道聚合、TrainMetric、DPOMetric 的 ragged ref_logps、EmbeddingMetric、GRPO old-logps 行数不匹配、上报辅助函数。
• test_ppo_metric.py（16 → 18）：PPOValueMetric 的 explained variance / clip / advantage 统计，以及 PPOMetric 与 GRPOMetric 的一致性。
• test_reporting.py（4 → 15）：JSONL 追加写与 flush、超时与后端失败不致命、SwanLab 前缀。
• test_throughput.py（5 → 9）：变长 batch 与 padded width、非法输入报错、窗口吞吐与显存归约。
• test_types.py（新增 3 个）：MetricRecord 不可变与 stage/status 校验。

同时包含两处配套的源码修复：

• GRPOMetric 在跳过 microbatch 时游标不前进：被丢弃的 microbatch 仍占用 old / adv / sampler 展平列表中的行，此前会让后续 microbatch 对齐到错误的 old logps，并使高 KL 记录中的全局样本序号偏移；空 logps 也会因 len(None) 抛 TypeError。现在两种情况都先按该 microbatch 的 logps 行数推进 cursor / _gsi_cursor 再跳过，空 logps 直接返回。
• TrainMetric 在计算上报 speed 时对零长度时间窗做保护：Windows 等平台的时钟粒度较粗，连续两次取时间可能得到相同值，step 增量 / 0 会抛 ZeroDivisionError 使指标计算中断；现在时间窗为 0 时 speed 记为 0.0，iters 与耗时不受影响。