Make Executor ReAct loop append-only for prompt-prefix caching

[yzx9]

Rewrite Executor.run to seed [system, user(task)] once and append one assistant turn (via the shared reconstruct_react_text) plus one user observation turn per step, instead of rebuilding a single growing user message each iteration. The static system prompt and all prior turns now form a byte-stable prefix, so provider-side prefix caching hits on every iteration. Previously only the system block cached; the whole rebuilt user message (Previous-Steps JSON + Iteration counter + last response) was a cache miss every step, with hit rate declining as the loop grew.

Folded correctness fixes (surfaced in plan-eng-review + /review):

• _generate_final_summary: render the string action_summary directly instead of calling .get('name') on it (AttributeError on exhaustion).
• tool_retry_counter: reset at run() start. Executors are reused across delegated tasks, so the counter leaked and could poison later runs.
• max_syntax_errors=3 budget with reset-on-success: abort a model stuck emitting malformed ReAct instead of spamming up to 15 correction turns. The budget resets after any cleanly-parsed turn, so only 3 CONSECUTIVE bad turns abort (not 3 slips scattered across a long run).

Extract reconstruct_react_text into copilotj/multiagent/react_format.py (LeaderAgent and Executor both import it; leader_multiagent imports Executor, so it could not live there without a circular import). Document in _build_enhanced_system_prompt that the tool text is load-bearing: the ReAct wrapper strips the tools= API param before the provider call, so the system-prompt text is the model's only view of the tools.

Add copilotj/test/multiagent/test_executor.py (9 cases incl. a post-format prefix-stability cache-contract test that runs messages through the Anthropic formatter rather than asserting raw list-append). Full suite 206 passed; ruff clean.

Assisted-by: Claude-Code:GLM-5.2

[00JackLu]

The code looks good, but I haven’t tested it.