fix(persistence): persist provider in model_config on model switch (#79536)

[RelaxJonh]

Problem

When a session falls back to a model on a different provider and the user switches to it with (or the Desktop model selector), writes only the model name to the table — the provider is lost. On the next session resume, the runtime recombines the persisted model with the primary provider from , producing a guaranteed on every turn.

Root Cause

sets but never writes or keys into the JSON blob. The session-resume code has no way to recover the correct provider/model pair.

Fix

Add an optional keyword argument to . When given, the provider and model are persisted into via , so session resume recombines the correct pair.

Files Changed

File	Change
	Add kwarg to , persist via
	Pass at both call sites
	Pass at fallback call site
	Add 2 tests: provider persistence + backward compat

Backward Compatibility

• When is not passed (default ), behavior is identical to before — only is removed from .
• All existing tests pass unchanged.

Fixes #79536

[teknium1]

Thanks — reviewed against current origin/main. The write-side gap is real (update_session_model never writes provider into the model_config blob), but PR #79604 by @webtecnica was filed ~7 hours earlier, fixes the identical bug with near-identical SQL at the same two call sites, and is a superset (adds the gateway_runtime.provider legacy-row fallback on resume plus resume tests). Earliest-submitter rule applies, so we'll take #79604 as the landing vehicle.

One note for the record that applies to both PRs: the WHEN model_config IS NULL THEN NULL arm means sessions with a NULL blob (the common case) still lose the provider — that will need to be handled in the surviving PR before/at merge.

Closing as duplicate — thanks for the fix and for confirming the read-side loop.