harness: empty-completion retry + duplicate-send guard for the turn loop

[Alexsun1one]

Problem

Two intermittent failures in the responses turn loop, both with weak models behind chat adapters:

1. Empty completion — a round with no text and no tool call was treated as a real end of turn, so an adapter-woken user could get no reply.
2. Duplicate send — a weak model loops on send_adapter_message and delivers the same reply as several near-identical variants. The loop only ends when the model calls no tool, and with no round cap it never stops.

Change

runResponsesTurnLoop gains two bounded guards:

• Empty-completion retry — a round producing nothing is retried up to 2x without consuming the tool-round-trip budget.
• Duplicate-send guard — once a reply has gone out, a further send_adapter_message with no real tool call in between is skipped; a round that is nothing but duplicate sends ends the turn. Any real tool call resets the state, so a genuine "acknowledge -> do work -> report result" sequence still sends twice.

Tests

shouldRetryEmptyCompletion is a pure function with unit tests. A loop integration test drives a fake runtime and asserts (a) a repeat send with no work between is skipped and the turn ends, and (b) a send after real work is allowed. 128 tests pass. runResponsesTurnLoop is exported for the test.

Harness-side turn hardening — validated in a live Lark adapter deployment where the duplicate-send guard demonstrably stopped a weak model from blasting the same reply multiple times.