R8a: make reasoning a graded level (off/low/medium/high) across all 5 providers

[dovvnloading]

Problem

The reasoning control was a bare bool (Thinking/Quick), applied only to Ollama/Llama.cpp models whose name happened to contain "qwen3" or "deepseek" - and completely absent for every cloud provider, since the composer hid the control entirely outside local mode.

Research

Looked at how Ollama and each cloud provider actually expose adjustable reasoning effort before touching any code - the mechanisms are genuinely different per provider and per model generation, not one shared knob:

• OpenAI: reasoning_effort (minimal/low/medium/high/xhigh, model-dependent)
• Anthropic: thinking.budget_tokens on older models; a newer effort param on Opus 4.7+ that rejects budget_tokens outright (a real 400 error) - mutually exclusive request shapes, not two names for the same thing
• Gemini: thinkingBudget (integer, 2.5-series) vs thinkingLevel (string, Gemini 3) - same story, different generations
• Ollama: think is a bool for qwen3/deepseek/qwq, but a required string level for gpt-oss - two different mechanisms on the same provider

Change

One shared off/low/medium/high vocabulary now drives all five, each mapped to whatever that provider/model actually supports - including being honest where a model can't truly grade: Ollama/Llama.cpp's bool-only models (qwen3/deepseek/qwq) get a real second axis via a plain-language reasoning-budget system-prompt hint layered on top of the bool, documented as prompt-based guidance rather than a fake graded API param.

The composer's reasoning control now shows real capability detection (does the currently resolved model support reasoning at all) instead of "is this a local provider" - cloud providers get the control whenever their active model genuinely supports it. Settings > Ollama/Llama.cpp pages get the same 4-option radio group the composer uses; adding equivalent Settings pages for the three cloud providers is left for a follow-up, since the composer's own popover already reads/writes their persisted level in the meantime.

Existing session.dat files migrate the old 2-value mode to the new scale (Quick → off, Thinking → high) rather than resetting silently; the three new cloud fields default to off (extended thinking on a paid API is an opt-in cost/latency tradeoff, never a silent default) while the local providers keep defaulting to high (local compute is free to the user).

Test plan

• Backend: 1021 passed, including a new dedicated test file for every provider's mapping function (25 tests) plus updated coverage across composer/settings
• Frontend: tsc --noEmit clean, lint 0 errors, 794 tests passed, build succeeds
• Verified live against the real, already-persisted session.dat: the migration correctly produced ollama_reasoning_level=high/llama_cpp_reasoning_level=high (from the prior Thinking value) and all three new cloud fields at off
• Verified live in the browser: the composer's Reasoning popover renders all four real options against the live backend, correctly shows "High" as active, and both Settings > Ollama and Settings > Llama.cpp pages show the same 4-option radio group in sync with the composer