Claude Auto fallback followups: PR #15687
Current head aa92c5d6aeb44a9e89dd393ec46fdb1e49c68b05
Pinned base1451e7cb23034d29774338608368a15d8845df36

Baseline merged suite:287 runnable cases. Final:288 passed in existing ClaudeAdapterV2.test.ts.
Existing fallback regression proves requestedAuto/accepteddefault same-model followup reuses process without retryingAuto.
Exact main production:1 target failed,1 passed(plan restoration),285 skipped.
Failure: expected ['auto'] to deeply equal []

R1 review found firstaccepteddefault baseline could strand a newlyAuto-capable model inManual after live model switch.
Added one parameter row in same existing fallback test.
Before model-switch fix:1passed(same model),1failed(changedAPI model),286skipped.
After fix:2passed,286skipped. Full288passed.
ChangedAPI afterfallback now closes old query and resumes same native session requestingAuto.
Fresh init negotiates new baseline. Existing backgroundwork replacement guard applies.
LaterPlan init/status still cannot replace firstaccepted baseline.

Two-file lint/format and server typecheck exit0. One existing unusedlayer lint warning; typecheck Effect suggestions only.
No UI layout changes; screenshots would not demonstrate this protocol behavior.
Proof uses actual adapter with injectedSDK stream/session. No fresh realClaude/account request or delegatedparent/child run.
macOS/Windows/iOS/Android/Electron/remote/relay live behavior not exercised.