Refactor: extract common base for purple-agent adapters (CAR + entropic)

[bordeauxred]

Why

Two purple-agent server implementations duplicate substantial scaffolding:

• `clawloop/environments/_car_purple.py` (344 lines)
• `clawloop/environments/_entropic_purple.py` (378 lines)

Both files implement near-identical versions of:

• `_convert_tools_to_openai()`
• `_normalize_assistant_msg()`
• session-state management
• harness-update protocol

A skilled engineer reading both files will spot the copy-paste immediately. It's also a maintenance liability: a fix in one is unlikely to be ported to the other.

What

• Create `clawloop/environments/_purple_base.py` containing `_PurpleAgentBase` with the shared methods.
• Refactor `_car_purple.py` to inherit from `_PurpleAgentBase` and override only the CAR-specific scenario handling.
• Refactor `_entropic_purple.py` to inherit and override only the CRM prompt parsing.
• All existing tests pass with no behavioral change.

Acceptance

• Total LOC across the three files reduced by ≥30%
• `grep "_convert_tools_to_openai" clawloop/environments/` returns one definition
• `pytest tests/environments/` green

Effort

~45 min. Bounded, low-risk, well-scoped — solid first-contribution issue.

Source

Code audit, 2026-04-18 (S1 finding).