fix(smith): use Ollama runtime context window

[edwin-zvs]

Summary

• preload Ollama models without inference and read the effective loaded context from /api/ps
• include Smith fixed prompt/tool overhead in pruning, probing, and auto-compaction budgets
• defensively cap malformed TUI context percentages at 100%
• document the Ollama runtime-window requirement in spec 0104

Verification

• cargo test -p construct-adapter-smith (226 passed)
• cargo test -p construct-cli modeline_context_usage_shows_full_token_text_over_a_bar
• cargo build
• isolated live ollama:qwen3.8 run reported context 20118/32768 (previously 20106/8000)

No recording: the behavioral proof is the emitted context event; the UI rendering itself is unchanged apart from displaying the corrected denominator and percentage.