fix: keep insights model metadata in sync after /model switch

[LeonSGP43]

Summary

• overwrite persisted session runtime metadata when an active session explicitly switches models
• keep /insights aligned with the live provider/base URL instead of preserving stale first-write metadata forever
• add a regression test that reproduces the stale Gemini-vs-GLM provider report and asserts the corrected insights output

Testing

• ./scripts/run_tests.sh tests/run_agent/test_switch_model_context.py tests/run_agent/test_switch_model_fallback_prune.py -q
• git diff --check
• uv sync --frozen --extra all
• uv run --frozen ruff check .
• repo lint-diff semantics (ruff clean; ty advisory signal recorded in proof)
• clean env -i pytest attribution against latest origin/main (30 head failures, 27 reproduced on base, remaining 3 isolated and passed on retry/base)

Fixes #3988.

[teknium1]

Thanks for the focused regression case. This is now implemented on current main by a later, stronger attribution design. Automated hermes-sweeper review confirmed that the reported /insights behavior is covered.

• hermes_state.py:2593-2660 records each API-call usage delta under the live model and billing route in session_model_usage, rather than rewriting a single mixed-session summary row.
• agent/insights.py:524-534 and agent/insights.py:578-647 aggregate those per-model rows, so a /model switch reports usage for each model that actually processed it.
• tests/agent/test_insights.py:323-350 verifies a DeepSeek-to-Claude mid-session switch produces separate correct totals.
• This shipped in cb7f6bbb2e5696bae2c595ca8a3a52cefb6a4a7e (feat(agent): track per-model token usage for mid-session model switches), which is an ancestor of current main.

The PR's one-row overwrite would instead collapse pre- and post-switch usage into the latest model, while current main preserves both routes.