Thesis

DeepSeek V4-Pro requires 27% of V3.2's per-token FLOPs and 10% of the KV cache at 1M context due to Hybrid Attention (CSA + HCA). The architecture only pays off in the long-context regime; aggressive routine compaction at 50K tokens defeats it. Make cache-maximal context the V4 default and treat compaction as an emergency mechanism, not routine hygiene.

Promotes #528 (currently drafted as opt-in) to default-on for V4 specifically. Legacy / V3 / custom-provider paths keep existing thresholds — raising compaction limits there could blow user bills.

Current behavior

• crates/tui/src/compaction.rs:38-39 — enabled: true, token_threshold: 50000, message_threshold: 50 regardless of model.
• crates/tui/src/cycle_manager.rs:65-68 — already V4-aware: hard cycle threshold bumped to 768K (~75% of 1M window). Compaction thresholds did NOT get the same treatment.
• The mismatch: cycle_manager rotates at 768K, but compaction.rs summarizes at 50K. We're paying for V4's cheap long context and then immediately throwing it away.

Proposed change

• Make CompactionConfig model-aware. When the active model is V4 family:
  • Raise token_threshold to ~500K (leaving ~500K headroom for: conversation, system prompt + tools, thinking-mode reserve, tool result accumulation).
  • Raise message_threshold proportionally or remove it.
  • Treat compaction as a near-cycle-boundary emergency, not routine.
• Promote Cache-maximal context mode: re-read active files instead of summarizing #528 ("re-read active files instead of summarizing") to default-on for V4. Working set stays resident; CSA/HCA handles the distant tail.
• V3 / legacy / custom-provider paths unchanged. Detect by model name prefix.

Open questions / risks

• Custom base URL serving a non-V4 model under a V4-prefix name → silently raised thresholds → blown bill. Need robust detection, opt-out config, or per-model override ([cycle.per_model] already exists in config.toml — extend the same pattern to compaction).
• 500K is a starting estimate; needs measurement against real long-running V4 sessions before being declared the default.
• Interaction with cycle_manager's 768K hard cycle: compaction at 500K and cycle at 768K leaves 268K of untouched-but-still-V4-cheap context. Confirm that's intentional.

Acceptance signals

• V4 sessions routinely hit 200K+ tokens without plan_compaction firing.
• working_set files stay resident across long turn loops.
• V3 sessions show no behavior change.
• A measurable cache hit rate improvement on long V4 sessions (cache stays warm because the prefix doesn't get rewritten).