Fix stuck model switches from false end-failure in session takeover - #1308

Nice work. Pulling the takeover decision out into planModelLockedSwitch in freebuff-session-api.ts is the right move — it turns an untestable inline conditional in use-freebuff-session.ts into a pure function with six focused unit tests covering the actual failure modes (ended-but-in-grace, ended-past-grace, swept row, mismatched model, unreadable row). That's exactly the kind of change that's easy to port by hand into a private tree, which is what this mirror needs.

The logic itself holds up: treating ended rows with an instanceId as still holding a slot (matching what holdsLiveFreebuffSlot already does elsewhere) closes a real gap, and refusing to delete a row for a different model preserves the existing safety property. The one-shot retry via the consumed explicit-pick marker is a sensible way to avoid a loop when the lock races and releases itself before the follow-up GET lands.

One gap: there's no test exercising the hook branch itself — the lockRaced retry path, the one-retry-then-revert behavior, and the message-only-on-explain claim from the PR description are all asserted by reasoning about the diff rather than by a test against use-freebuff-session.ts. Given how async and store-coupled that file is, I understand why you stopped at the planner boundary, but a maintainer will likely want at least a smoke test on the hook's model_locked branch (mocking callFreebuffSession) before merging, since that's where the retry/no-loop guarantee actually lives, not in the pure function.

Good scope discipline otherwise — no touches to forbidden paths, no drive-by reformatting of the pre-existing prettier violations you called out.

Follow-up on the hook-branch ask: the branch is now production code with its collaborators injected, so it is testable without mounting the hook.

runModelLockedTakeover(explicitPickModel, lockedModel, deps) (cli/src/hooks/use-freebuff-session.ts) takes the whole model_locked branch — deliberate-pick guard, follow-up GET, planner call, DELETE, and both notices — and returns repick | revert | stale. tick keeps the four wiring lines that turn that outcome into nextMethod / schedule(0).

cli/src/hooks/__tests__/model-locked-takeover.test.ts (13 tests; no stores, timers, spies, or mock.module) covers the three points:

• lockRaced: {status:'none'} and {status:'ended'} with no instance id both return repick with zero notices and zero DELETEs; the active row and the grace-window ended row return repick only after releaseSlot receives exactly that row.
• one retry: the marker read-and-clear moved into takeFreebuffExplicitPick(), so the retry test drives rounds through production consumption rather than a test-side simulation — ['repick','revert'], one follow-up GET, no notice. Dropping the pendingExplicitPickModel = null line turns both that test and the direct accessor test red.
• message only on a real attempt: both notices are asserted as full strings with both display names in their slots (swapping labels.current / labels.requested in one template goes red), and the failure wording is asserted on exactly the four explain shapes — foreign-model row, superseded, throwing read, refused DELETE — with silence asserted on repick, stale, background rejoin, and a pick naming the locked model.

Gates, repo-pinned bun 1.3.14: tsc --noEmit -p cli exit 0. Full cli suite 3065 pass / 18 fail at head vs 3046 pass / 18 fail at base 3f00c772a, failing set identical — 16 packages/internal/src/env import errors (paths absent from the snapshot) plus 2 pre-existing /reasoning reds; the delta is the 19 added tests. QueuePanel's two render tests flake (1 red in 5 isolated reruns at base), which is why that count moves run to run. Twelve single-line mutations of the branch, the planner, and the marker clear each turn the suite red; all reverted.

Two known limits, stated rather than papered over: the four wiring lines in tick are still asserted by reading — reaching them needs the hook mounted (React, fake timers, mock.module() on the session client), which the DI rule rules out, and they are byte-identical to what you reviewed in the first commit — and startFreebuffSession is not used to arm the marker in tests because it persists the model preference to disk. Branch squashed to aa465f973 and rebased onto 3f00c772a.