Add default-off classify early stopping

[scrimshawlife-ctrl]

Summary

Optional, default-off early stopping for the classify training loop. HYPERLEX_TRAIN_EPOCHS stays the max-epoch cap and does not turn this on.

• Enable with HYPERLEX_EARLY_STOP=1, HYPERLEX_EARLY_STOP_PATIENCE, and HYPERLEX_EARLY_STOP_MIN_EPOCHS.
• Supported only when HLX_SELECT_METRIC=classify_macro_f1_nonnone. Negative patience, invalid minimum epochs, and other selection metrics fail before the optimizer is constructed.
• After a scored epoch records any strict improvement, stop when epoch_index - best_epoch >= patience and at least minimum_epochs epochs have been scored. Ties keep the earlier checkpoint and consume patience.
• A best at epoch 3 with patience 4 and no later strict improvement stops after epoch 7 (8 epochs scored) when the cap is 12.
• The declared metric's best checkpoint is still restored after early stopping or after the cap.
• Every epoch-progress.jsonl row gains observational seconds, rounded with round() to 6 decimal places: epoch_wallclock_seconds and training_elapsed_seconds. They do not affect selection or stopping.
• A completed train-receipt.json records stop_reason (max_epochs or early_stopping) and training_elapsed_seconds. A failed loop still raises and does not write that completion receipt.

Tests

Focused tests: tests/shadow/test_hyperlexical_early_stop.py (24 passed) on Python 3.10.21, 3.11.16, and 3.12.3, using an injected monotonic clock and no sleeps.

Full HYPERLEX_OFFLINE=1 HYPERLEX_NO_RATE_LIMIT=1 PYTHONPATH=src python3 -m pytest -q:

• Python 3.10.21: 1039 passed, 13 skipped, 16 subtests passed
• Python 3.11.16: 1039 passed, 13 skipped, 16 subtests passed
• Python 3.12.3: 1047 passed, 7 skipped, 16 subtests passed

This does not seal an experiment, allocate a reserve, launch training, score a checkpoint, or move BEST. Do not merge from this description alone.