fix: sync model.context_length on --global model switch

[solink7]

Problem

When using /model <name> --global, the resolved context length was displayed correctly in the switch confirmation but not persisted to config.yaml. This caused:

• model.context_length stayed at the old model value
• Compression thresholds and token budgets used stale context window
• Users switching between models (e.g. gpt-5.5 to mimo-v2.5-pro) got incorrect context limits

Root Cause

In _apply_model_switch_result(), the context length was resolved via resolve_display_context_length() for display, but the persist_global block only saved model.default and model.provider — it never saved model.context_length.

Fix

• Extract ctx variable to outer scope so it is available after the try/except block
• When persist_global is true and ctx is resolved, save it with save_config_value("model.context_length", ctx)

Testing

• All existing model switch tests pass (28/28)
• Full tests/hermes_cli/ suite: 3168 passed, 13 failed (pre-existing, unrelated)

[alt-glitch]

Related to #10157 and #10366 (warn on stale context_length), and #15706 / #11437 (context-length alignment). This PR goes further by auto-syncing the value rather than just warning.

[solink7]

Superseded by #17246 which includes both the config persistence fix and runtime display update.