Qt-removal R7.4c: Llama.cpp settings page + native file-picker capability

[dovvnloading]

Summary

• Closes R7.4 (all 3 of R2.5d's deferred settings pages are now real): adds backend/native_dialogs.py - the one genuinely new capability this increment needed - wrapping pywebview's create_file_dialog into pick_file/pick_folder, gracefully returning None when no native window exists rather than raising.
• backend/settings.py gains the Llama.cpp page for real: reasoning mode, runtime tunables (chat format/n_ctx/n_gpu_layers/n_threads), GGUF model scan (system + real folder-picker), and staged-then-Saved chat/title model paths (saveLlamaCppSettings validates then persists, matching legacy's own two-phase design) - 12 new intents, 11 new payload fields.
• Building the picker capability also retroactively un-defers the Ollama page's own long-disabled "Scan Folder..." button (pickOllamaScanFolder), the exact dependency R7.4b's own docstring flagged.
• A 5-way adversarial review (backend/frontend/integration/legacy-parity/security - the 5th lens added since this is the first native-OS-integration point in the whole rewrite) found and fixed 4 real issues, all mutation-tested: a stranded reentrancy gate if the native dialog itself raises; an unlocked read race in the live-reapply paths; 5 collapsed legacy error messages restored verbatim; missing initial-directory seeding + a dropped "All Files" filter fallback on every native dialog call site, both restored to match legacy. Security review found nothing above informational (single-user localhost desktop app; GGUF loading is always lazy/non-preloading; no path-traversal-shaped concatenation anywhere).
• One narrow, already-accepted-elsewhere risk documented inline rather than chased further: the new numeric-field/chat-format draft buffers (added to fix a real, deterministic multi-digit-typing bug their own test caught) can still have a same-session echo of an older keystroke stomp a newer draft under sustained fast typing during a slow round trip - matches OllamaTaskField's own already-accepted unbuffered risk, not a new regression.

Test plan

• python -m compileall -q . clean
• Burn-down gate (tests/test_no_qt_anywhere.py) unchanged: 152/84/68
• Full backend pytest: 2496 passed (2494 + 2 net new)
• Full frontend npm run check (schema-check/tsc/eslint/vitest/build): 1100 vitest passed, 0 lint errors, build clean
• All 4 fix-pass regressions mutation-verified (revert the fix → test fails), including both new native-dialog-failure tests
• Live WS drive against a real running backend exercising the native dialog's graceful no-window path and a real, unmocked GGUF filesystem scan, using an isolated scratch settings file - the real ~/.graphlink/session.dat confirmed byte-identical before and after