fix(insights): persist token usage for non-CLI sessions

[kshitijk4poor]

Problem

/insights can show 0 tokens for non-CLI sessions even when those sessions clearly consumed model usage.

I confirmed the split-brain persistence path described in the repro notes:

• run_agent.py accumulated token usage on every API call, but only wrote it to SQLite when platform == 'cli'
• gateway sessions also attempted a second higher-level token/model write after run_conversation() returned
• cron and other non-CLI/non-gateway sessions could miss persistence entirely, while gateway accounting stayed split across two code paths

That left /insights vulnerable to missing token/accounting data for Telegram/cron rows.

Root Cause

Token persistence was split across two places:

1. run_agent.py

   • had the real per-API-call usage data
   • but skipped non-CLI sessions because of a CLI-only guard

2. gateway/run.py

   • did a second session-store write after the run finished
   • duplicated token/model persistence responsibilities instead of keeping them in one place

Fix

• removed the CLI-only guard in run_agent.py
• now any session with a session DB + session ID persists token usage directly from the agent loop
• removed the gateway's redundant token/model/cost write in gateway/run.py
• simplified gateway.session.SessionStore.update_session() so it now only updates lightweight session metadata (updated_at + last_prompt_tokens) used for context-window tracking / compression decisions
• removed stale gateway-session test coverage for the deleted write-through path
• added regression coverage for Telegram and cron session persistence
• updated gateway message-handling assertions to match the consolidated write path

Testing

• source /Users/kshitij/Projects/hermes-agent/.venv/bin/activate && python -m pytest tests/test_token_persistence_non_cli.py -q
• source /Users/kshitij/Projects/hermes-agent/.venv/bin/activate && python -m pytest tests/test_token_persistence_non_cli.py tests/gateway/test_status_command.py tests/gateway/test_session.py tests/test_insights.py -q

Notes

This now fully follows the hybrid approach from the repro notes:

• agent-level persistence for all platforms
• gateway-level last_prompt_tokens only
• no dead gateway token/model persistence path left behind