condenser hard context reset: preserve the leading system prompt instead of summarizing it at offset 0

Problem

LLMSummarizingCondenser.hard_context_reset() in openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py summarizes all events in the view:

return self._generate_condensation(
    forgotten_events=view.events,   # ALL events, including the system prompt
    summary_offset=0,
    max_event_str_length=max_event_str_length,
)

with summary_offset=0. The resulting CondensationSummaryEvent (a user message, see the related edge-case issue) is inserted at index 0 — before (or replacing) the SystemPromptEvent. Combined with keep_first semantics, this means the system prompt can be entirely forgotten and the re-injected conversation opens with a user-role summary with no system message.

This violates the repo convention that every LLM request starts with a system message (see parent tracking issue #5144). It is an edge case triggered by LLM-error/hard-reset recovery, but the resulting view violates the invariant that init_state otherwise guarantees.

Proposed change

In hard_context_reset, exclude the leading SystemPromptEvent from the forgotten events (or always preserve it in the retained prefix):

1. Keep the SystemPromptEvent (and any essential prefix events) out of forgotten_events, e.g. pass view.events[max(0, first_non_system_index):] and compute a summary_offset that places the summary after the preserved system prompt.
2. Ensure both the sync hard_context_reset and async ahard_context_reset paths do the same.

Add a regression test: trigger a hard context reset and assert the resulting view/messages still begin with a system message.

Note this issue and the keep_first edge-case issue are related but distinct: this one concerns the hard-reset path that summarizes the entire view including the system prompt with summary_offset=0.

Parent issue: #5144

This issue was created by an AI agent (OpenHands) on behalf of juanmichelini.

Start in openhands-sdk/openhands/sdk/context/condenser/llm_summarizing_condenser.py and inspect the sync hard_context_reset and async ahard_context_reset entry points. Add a regression test that triggers a hard context reset and verifies the resulting view and messages still begin with a system message. Confirm both reset paths preserve the leading SystemPromptEvent while summarizing the remaining events.