feat(v3.05.82): adaptive Markdown streaming, visual /context, deepsee…

[chauncygu]

…k-v4-flash 1M

Adaptive per-device Markdown streaming (ui/render.py):

• Auto-select a streaming tier in auto_stream_mode(): "live" (in-place Rich redraw on capable terminals incl. modern SSH emulators), "commit" (append-only progressive Markdown — completed blocks render and print permanently with zero cursor movement, so frames can never duplicate over SSH / Apple Terminal / pipes / with wide CJK-emoji text), and "plain" (raw tokens, no Rich).
• Override via /config stream_mode=live|commit|plain (legacy rich_live still works).
• Replaces the old crude "disable Live whenever SSH" gate that dropped SSH users to raw tokens. Adds tests/test_stream_modes.py (26 cases incl. a regression that commit mode emits zero cursor sequences even on a TTY with CJK text).

Visual /context grid (commands/core.py):

• 20x10 colored cell grid of context-window usage by category (system prompt / tools / memory / skills / messages / free space) with per-category tokens + %, adapting to the model's real window; ASCII (#/.) fallback on non-UTF-8 terminals.

deepseek-v4-flash 1M context window (providers.py):

• Registered in _MODEL_CONTEXT_LIMITS so the prompt %, /context, and compaction trigger reflect the true 1M window (deepseek-chat / v4-pro stay at 128K).

UI: add a blank line above the startup logo so it sits farther below the prompt.

Docs: README contributors update; one-line README news + detailed docs/news.md entry; features.md / reference.md updated; version bump to 3.05.82.