fix: clear ended_at metadata when messages are appended to an ended session

[nicha16]

Problem

append_message() in the SessionDB increments message_count on the sessions row but never clears ended_at or end_reason. When a session is ended (e.g. by daily reset or idle timeout) and later receives new messages, the stale ended_at causes hermes sessions list to sort the session by its old end time rather than the actual last activity.

This produces incorrect ranking: the session can appear at position #29 instead of its true rank (~#12) by last-message timestamp.

Fix

Add ended_at = NULL, end_reason = NULL to the existing UPDATE in append_message(), in both the tool-call and non-tool-call branches. This is free — the row is already being written. For the 99.9% of sessions where ended_at is already NULL, the SET NULL is a no-op. It only affects the edge case where messages arrive for a previously-ended session.

Test

Adjusted the compression-tip ordering test so the root message is appended before the root is ended (realistic order), matching the fix's assumption that append_message clears ended_at. All 288 tests in test_hermes_state.py pass.

Implementation

# Before
UPDATE sessions SET message_count = message_count + 1 WHERE id = ?

# After
UPDATE sessions SET message_count = message_count + 1,
                    ended_at = NULL, end_reason = NULL WHERE id = ?