Refactor: extract helpers from learning_loop() god-function

[bordeauxred]

Why

`clawloop/core/loop.py:156` defines `learning_loop()`, a single function ~420 lines long that handles:

• task sampling
• batch episode execution
• archive writes
• evolver context construction
• per-layer transactions
• rollback on failure
• generation flush
• structured logging

Reading the loop end-to-end requires holding all seven concerns in your head at once. New contributors trying to add a feature touch this function and can't tell what's load-bearing vs incidental.

What

• Extract `EpisodeCollectorRunner` — handles task sampling + batch execution (lines ~256-380). Lives in `clawloop/core/runner.py`.
• Extract `LayerTransaction` — handles snapshot → forward_backward → optim_step → rollback-on-failure protocol (lines ~410-510). Lives in `clawloop/core/transaction.py`.
• Extract `ArchiveRecorder` — handles RunRecord/IterationRecord/EpisodeRecord writes (lines ~520-573). Lives in `clawloop/core/archive_recorder.py`.
• `learning_loop()` becomes a thin orchestrator: ≤80 lines, reads top-to-bottom as the algorithm.

Acceptance

• `learning_loop()` is ≤100 lines
• All existing tests pass with no behavioral change
• New unit tests for each extracted helper (small, focused)

Effort

~3 hours. Touches a hot path so testing matters; existing 1023-test suite gives good coverage.

Source

Codex code audit, 2026-04-18 (S1 finding).