Put the local-model daemon back on its defaults for an untuned assessment

[atomantic]

Summary

An untuned assessment only applied its launch tuning when the run had knobs. So a daemon the previous run tuned kept serving under that tuning while the reading was stored with tuningKey: '' and rendered as Backend defaults — and compareTunings ranked every real tuning against it, making the one row a user reads as the baseline the least trustworthy row in the table. #4775 made an empty knob set meaningful for llama.cpp behind reset, but an ordinary Measure still took the no-applier branch on every runtime.

An empty knob set is now a real instruction wherever the manager can honour it (resetsOnEmpty), for a plain Measure as well as a sweep baseline:

• llama.cpp — restores the launch line PortOS displaced when it first tuned the daemon. Deliberately not CLEARED_TUNING: that clears sweepable knobs the user may have set on the LLMs page, which only a sweep — having captured them — is entitled to do. Restoring the captured line undoes what PortOS applied and nothing else, so reset stays opt-in and sweep-only exactly as documented.
• Ollama — actively unsets what it exported. launchctl setenv writes into a launchd domain that outlives the daemon, so omitting a variable from the next restart does not clear it. The baseline is the env in effect before the tuning, not an empty one: ensureContextWindow puts the user's agent context window on the same OLLAMA_CONTEXT_LENGTH a tuning uses, and stripping it would undo a setting they chose.
• LM Studio — reloads through lms load with no flags, tracking per model which loads carried them so an already-default model is not cold-loaded for nothing.
• MTPLX — resetsOnEmpty: false. Its manager refuses an empty knob set rather than relaunching without flags, so an untuned run must not claim it reset anything.

This closes #4763 as well: its three scope items are the Ollama and LM Studio resets and the flag flip.

tuningApplied gains a third state at the manager boundary: null means nothing needed to change — the honest answer for a run with nothing to apply, and what keeps untuned readings in the recommendations rather than dropping them out via a false (the regression #4763 warned about). A reset that fails records false with its reason, and both the server's exclusion sentence and the client chip word that case for what it is — the daemon could not be put back on defaults — instead of "tuning was not applied", which contradicts the row's own "backend defaults" label.

Test plan

• localModelAssessments.test.js — every applier that can reset is asked to on an untuned run; a runtime that cannot still claims nothing; a failed reset records false with its reason; llama is asked with reset: false so an ordinary Measure never clears the user's flags; the exclusion reason is worded for which direction failed.
• llamaServerManager.test.js — restores the pre-tuning launch line; relaunches nothing when already at baseline; keeps the original baseline across successive tunings; drops it when the daemon is stopped, or when a fresh server is started over a crashed one.
• ollamaManager.test.js — restarts without the env it applied; unsets launchd variables (including ones the next tuning stops naming, and ones exported before a failed restart); restores the pre-tuning env rather than stripping it; drops the baseline however the daemon was stopped; does not claim a variable whose setenv failed.
• lmStudioManager.test.js — reloads without flags; no-ops when already default; refuses when a resident model would not unload; loads a non-resident model without reading its unload error as a refusal; attempts the unload when the loaded-model list could not be trusted.
• assessmentTuningNotice.test.js plus both component suites for the reworded rows.
• Full suites green: server 32600 passed, client 9233 passed, lint clean.

Notes

Reviewed locally over six rounds (codex ×3, claude ×3) before the PR opened; 15 findings, 14 fixed. Every guard was mutation-tested — each fix reverted in turn to confirm the corresponding test fails.

One finding was deferred rather than fixed here: getLlamaServerStatus collapses "PM2 unreadable" into "not managed", so a transient read failure can drop a model's baseline from the ranked list. It needs a change to that function's status contract, and the refusal itself is the safe direction. Filed as #4780. (mtplxServerManager already distinguishes the two — that is the pattern to follow.)

Closes #4759
Closes #4763