Review + reproduction evidence

Pinned review: 7fa1f1ffc8 (PR head b8dcb645d8 plus two fixes I pushed). Three findings on the new controls, all reproduced by hand and fixed in 36dcad8366 and 7fa1f1ffc8.

I ran the flow on fresh upstream main (496acd8f) and on main plus this branch, installed with install.sh --local --no-torch on Modal (CPU, llama.cpp), Chromium 1440x900, unsloth/Qwen3-0.6B-GGUF Q4_K_M and unsloth/SmolLM2-135M-Instruct-GGUF Q4_K_M:

• main: the API page toolbar has Pause, Unload, Refresh, Clear log and API settings only. With no model loaded there is no way to load one from this page.
• PR head, core flow works: pick Qwen3 Q4_K_M in the new picker, the monitor reports unsloth/Qwen3-0.6B-GGUF:Q4_K_M · 40,960 ctx; Unload gives None; Reload (tooltip Reload unsloth/Qwen3-0.6B-GGUF · Q4_K_M (no model loaded)) brings the same quant back.

Findings on the PR head, each reproduced with the same steps:

1. The picker eject unloads the wrong model after an API auto-switch. With "Switch model by request" on, load Qwen3 from the picker, then an API client requests SmolLM2. The trigger shows SmolLM2, but its eject called ejectModel(), which unloads Chat's stale selection. Server log: Unloaded model: unsloth/Qwen3-0.6B-GGUF (a 200 no-op), and /api/inference/status still reported active_model: unsloth/SmolLM2-135M-Instruct-GGUF 15 s later. The trigger eject now goes through the page's existing unloadResident path, the same one the Unload button uses. After the fix, the same steps leave nothing loaded.
2. The per-row Eject unloads the wrong model with several models kept loaded. With "Keep multiple models loaded" on and Qwen3 plus SmolLM2 resident, clicking Eject unsloth/Qwen3-0.6B-GGUF unloaded SmolLM2 (Unloaded GGUF model: .../models--unsloth--SmolLM2-135M-Instruct-GGUF/...) and left Qwen3 loaded. The wrapper dropped the row's id. It now forwards it, and the same click unloads Qwen3 (Unloaded GGUF model: unsloth/Qwen3-0.6B-GGUF) and keeps SmolLM2.
3. The trigger showed the quant twice and could show a stale one. It read Qwen3-0.6B-GGUF:Q4_K_M GGUF · Q4_K_M. After an API switch to Q8_0 it read Qwen3-0.6B-GGUF:Q8_0 GGUF · Q4_K_M, because the chip came from Chat's store. The id and the quant are now both taken from the monitor's <id>:<quant>, so it reads Qwen3-0.6B-GGUF GGUF · Q4_K_M and the loaded row is matched in the list.

I also made the load call match Chat's picker: it now forwards loadId, isVision and downloadPresentation, and passes previousConfig, so a failed or cancelled switch rolls back to the resident model's settings. Reload passes previousConfig too.

• Typecheck passes on main plus this branch. eslint is clean on the new files. biome reports the same 15 diagnostics on api-monitor-page.tsx as on main.
• tests/api-monitor-reload-last-model.test.ts: 6/6 pass, including a new splitActiveModel case. On the Actions A/B mirrors the file passes on the PR side and fails with ERR_MODULE_NOT_FOUND with the source reverted. The two red jobs on both sides are not from this PR: locale parity on the settings.resources.*.modelsDisk keys, and the Windows system-status-verdict.test.ts "no tile fabricates a usage percentage" case.
• Frontend only, no backend change. The controls render only on the API page, not in the floating monitor overlay.

Two further review rounds on a mirror. The first raised three items: the duplicate quant, the dropped loadId and the missing previousConfig. All three are fixed above. The second, on 7fa1f1ffc8, raised two:

• Set throwOnError on the load calls: not taken. selectModel already reports failures itself (setModelsError and the error toast). With the flag, a pick replaced by a newer one would also throw "Model selection was superseded by a newer choice." and show up as a false inline error.
• The initial refresh({ includeLoras: false }) skips listLoras(). If someone opens /api-monitor before Chat has refreshed the store, trained LoRAs are missing from this picker. I could not reproduce it here (no trained LoRA on a CPU box), so I am leaving it as a suggestion:

    refresh();

(api-model-load-controls.tsx, the first line of the mount effect.)

Conclusion: the gap in #11189 is real on main, and the PR closes it. With the fixes above, load, unload, reload and both eject paths behave correctly on current main. Not covered by hand: an external provider selected in Chat (no provider key on the test box), and Windows/macOS beyond the Actions unit jobs.