This is a tight, well-tested fix. The per-model Record plus a localStorage layer cleanly separates the two symptoms (the cross-model leak vs the lack of cross-session memory), and it deliberately closes the stale-value shadowing trap (CRF-22) litigated on the superseded #27377 by validating the stored value against the model's current effort set before it enters the precedence chain. Test density is high (171 test / 65 production LOC) and the three new stories drive the real slider and the real submit payload rather than asserting mock echoes.

Severity count: 0 P0-P2, 2 P3, 1 P4, 1 Nit, 3 Notes.

The two P3s are worth a look before merge. CRF-2: the root-override effort is gated behind !hasValidUserSelectedModel, so selecting your override's own model in the dropdown drops the override effort back to the model default (verified: handleModelChange always sets the flag), which contradicts the PR's "the override still applies when nothing is persisted." CRF-3: effort persists on every slider move rather than on chat creation, so an unsubmitted preview durably outranks the account override next session, asymmetric with the sibling lastModelConfigID commit-on-create behavior. Both are bounded but deserve a deliberate call rather than silent acceptance.

Process note: the precedence and stale-value logic is covered only by Storybook play assertions, and the blocking CI job runs vitest run --project=unit only, so a regression in that chain produces no failing blocking-CI signal. This is a repo-wide infrastructure gap, not introduced here, and the author followed the project's story-based convention; flagged as CRF-1 so the coverage assumption is explicit.

Hisoka, on the diff: "I came looking for a fight. This one held its ground."

🤖 This review was automatically generated with Coder Agents.

Resolutions in 1ed7672:

• CRF-2: fixed. Dropped the !hasValidUserSelectedModel term; selectedModel === rootOverrideModelID already scopes the override and it sits last in the chain, so a per-model or persisted pick still wins. New story ManualReselectKeepsRootOverrideEffort pins it.
• CRF-3: deliberate keep of persist-on-change. It is what "remember the previously selected level" means to a user; persisting only at create would lose the choice when the create fails or the user navigates away. Description already says "selected", not "used".
• CRF-4: the meta-level beforeEach: localStorage.clear() (line 125) already runs for every story including SubmitsReasoningEffort, providing the invariant. No story-level clear added; a redundant one on the new story was also removed.
• CRF-5: declined. The unit test covers the no-throw contract; the in-memory path is simple and its failure mode benign. Not worth a story.
• CRF-1/6/7: noted, no change (repo-wide CI gap; bounded unpruned keys the reviewer called not worth machinery; single-candidate stale validation the reviewer rated low impact).