Summarize long sessions earlier, at a token ceiling #95

Problem

A session ran with a large context window and never summarized, so the prompt grew to about 84K tokens. Past about 75K, Qwen3.6 35B (3B active parameters) degraded in several ways:

• it looped on one command;
• it ignored a clearly printed fact from its own test output (grid_layout=3×3, when the user wanted 2×3);
• it lost track of what had already been tried.

lcode compacts at 85% of the window, which with a 256K window means about 220K tokens: far past where small models work well.

Proposal

• A token ceiling (setting compact_at, default around 64K tokens) on top of the percentage: lcode prunes, then summarizes, when either is reached.
• Keep the user's goal: the summary keeps the user's requests verbatim (the most recent ones at least), so the goal survives.
• Show it: /context shows the ceiling.
• Measure with lcode bench --session before choosing the default.

Acceptance criteria

• A configurable ceiling, with a measured default
• Summaries keep the user's requests
• Docs