Qt-removal R7.4b: Ollama settings page (backend+frontend)

[dovvnloading]

Summary

• Closes the second of R2.5d's 3 deferred settings pages: reasoning mode, per-task model assignment (auto/inherit/explicit), real system model scan, and real model pull - 4 new backend/settings.py intents + a new OllamaPage/OllamaTaskField pair in SettingsDialog.tsx.
• Proactively applies the R7.4a "fix classes not instances" lesson at design time: every intent argument is coerced before use, every read-modify-write happens inside one locked closure, and the reasoning-mode live-reapply collapses its check-then-act into a single hop so a concurrent provider switch can't be clobbered back to Ollama.
• A 4-way adversarial review (backend/frontend/integration/legacy-parity) found and fixed 3 real issues: a stranded reentrancy gate on a scan/pull persistence failure, a missing task_chat + "inherit" normalization gap, and a misleading legacy-parity comment (the reasoning-mode live-reapply is actually a fix for a real legacy staleness bug, not a port of existing behavior). All fixes mutation-tested against reverted code.
• Deliberately deferred, unchanged from the R7.4 scoping call: the Ollama page's "Scan Folder..." button stays disabled (visible tooltip, not silently broken) pending R7.4c's native folder-picker capability. "System Scan" is fully real.

Test plan

• python -m compileall -q . clean
• Burn-down gate (tests/test_no_qt_anywhere.py) unchanged: 152/84/68
• Full backend pytest: 2456 passed (2453 + 3 net new)
• Full frontend npm run check (schema-check/tsc/eslint/vitest/build): 1081 vitest passed, 0 lint errors, build clean
• All 3 fix-pass regression tests mutation-verified (revert the fix → test fails)
• Live WS drive against a real running backend with a real Ollama server (10 real models scanned, a real pull succeeded), using an isolated scratch settings file - the real ~/.graphlink/session.dat confirmed byte-identical before and after