Add configurable console metrics filter with presets

[YujiaBao]

Summary

Add a configurable filter system for the console metrics table (PrettyPrintLogger), with named presets that users can select via CLI or Config.

Depends on #522 (telemetry unification).

Motivation

The console table shows ~50 metrics per step. Most users only care about a handful (reward, progress, key timing). The rest is useful for debugging but clutters the default view. All metrics are always written to metrics.jsonl regardless of the console filter.

Changes

ConsoleMetricsFilter dataclass (ml_log.py)

Pattern-based filter using fnmatch globs:

• include: list of patterns — metric must match at least one to be shown
• exclude: list of patterns — metric is hidden if it matches any (takes precedence)

Three presets (CONSOLE_METRICS_PRESETS)

• "compact" (default for training Configs): Shows key metrics — reward, correct/format, rollout stats (turns, tokens), progress, optim, and top-level timing (time/total, time/sampling, time/train_step). ~16 rows instead of ~50.
• "detailed" (default for direct setup_logging() callers): Shows everything except :total/:count aggregates.
• "all": No filtering — shows every metric.

Config integration

New console_metrics: str = "compact" field on all three training Configs (RL, SL, DPO). Passed through to setup_logging().

Discoverability

On the first log_metrics call where metrics are filtered, a one-time hint is logged:

Console metrics using 'compact' preset (37 metrics hidden).
Set console_metrics='all' to show all, or 'detailed' for more.
All metrics are always written to metrics.jsonl.

CLI usage

console_metrics=compact       # curated summary (default)
console_metrics=detailed      # everything except :total/:count
console_metrics=all           # no filtering

Backward compatibility

• Direct setup_logging() callers (e.g., rl_loop.py, sl_loop.py) that don't pass console_metrics get "detailed" — same behavior as before.
• Training Config-based callers get "compact" — UX change (fewer metrics shown), but the hint message tells users how to see more.
• All metrics are always written to metrics.jsonl, Wandb, Neptune regardless of the console filter.

Before/after console output (RL sync, step 0)

Before (59 keys in metrics.jsonl, all shown in console):
39 time/* rows + 13 env/* rows + 7 other = ~59 rows

After (compact preset, 16 rows in console, 51 keys in metrics.jsonl):

│ env/all/ac_tokens_per_turn    │ 5.000000  │
│ env/all/correct               │ 0.000000  │
│ env/all/format                │ 0.000000  │
│ env/all/ob_tokens_per_turn    │ 38.500000 │
│ env/all/reward/total          │ -0.100000 │
│ env/all/turns_per_episode     │ 1.000000  │
│ optim/entropy                 │ 0.000774  │
│ optim/kl_sample_train_v1      │ 0.000025  │
│ optim/kl_sample_train_v2      │ 0.000000  │
│ optim/lr                      │ 0.000010  │
│ progress/batch                │ 0         │
│ progress/done_frac            │ 0.500000  │
│ time/run_evaluations_parallel │ 0.000016  │
│ time/sampling                 │ 6.675279  │
│ time/total                    │ 36.008096 │
│ time/train_step               │ 28.200690 │

Test plan

• All 100 unit tests pass
• ruff format, ruff check, pyright clean (0 new errors)
• test_recipe_math_rl::test_math_rl_sync e2e smoke test passes with compact preset
• Hint message displays correctly on first step
• All metrics still written to metrics.jsonl (verified: 51 keys)

🤖 Generated with Claude Code