Manual placement: log the --fit verdict the launch actually carries

[deepspace28]

Summary

Follow-up to my report in #10821. In Manual GPU-memory mode with an explicit layer count, the backend's GGUF decision line interpolated use_fit before the Manual branch turned it off — so every Manual load logged --fit: on while the launch it described six milliseconds later carried --fit off. Paired with the deliberately emptied GPU probe, the two lines read "failed GPU enumeration, proceeding unprotected" instead of "Manual mode, the user owns placement".

Fix

load_model decides Manual placement from a condition that is knowable well before the log: the function arguments gpu_memory_mode and gpu_layers. So the change:

1. Hoists the verdict to before the decision line, from the same predicate the launch branch applies (gpu_memory_mode == "manual" and gpu_layers >= 0), so the log can no longer disagree with the argv.
2. Labels a deliberately empty list — GPUs free: [] (manual placement) — so an intentional stand-down is distinguishable from a failed probe at a glance. The dedicated empty-probe warning (correctly gated on _detected_gpus) keeps catching the genuinely failed case.
3. Re-asserts use_fit = False in the Manual launch branch, so the placement try/except's --fit on fallback can never override Manual between the hoist and the launch.

A side effect worth noting: _placement_verdict_partial reads use_fit right after the log, so under the old order a Manual load recorded a partial-placement verdict ("the planner ran and could not fit"). With the hoist it reads the Manual override — False, which is the correct claim for a mode where nothing was fitted.

Tests

New studio/backend/tests/test_manual_mode_decision_log.py — source-level on load_model, since the log site cannot be driven from a unit test:

• Manual turns the fit off above the decision line (this trio fails on main with the order reverted)
• The hoist consults the same predicate the launch branch applies
• An empty probe can carry the (manual placement) label
• The launch branch still re-asserts use_fit = False before emitting --gpu-layers, so the placement fallback cannot leak into Manual
• The partial-placement verdict reads the post-hoist flag
• Manual + Auto layers (the fitter-carrying sibling branch) stays untouched

All 6 pass; test_empty_gpu_probe_reason.py and test_log_filter_no_truncation.py stay green (23 + 9).

Closes #10821

[deepspace28]

@pre-commit-ci[bot] re-run