Token audit: verify context triggers use active request size, not lifetime usage

[Hmbown]

Problem

A token-management system can make a strong model behave dumb if it triggers context rewrites from the wrong token basis.

In the v0.7.5 hard-cycle review, we found one concrete foot-gun: lifetime API usage can count the same stable prefix repeatedly across turns. Cycle/seam decisions should be based on the active request input estimate, with future assistant-output headroom considered against the model window, not cumulative billed/usage totals.

Audit scope

Read first:

• crates/tui/src/cycle_manager.rs
• crates/tui/src/core/engine.rs
• crates/tui/src/core/engine/turn_loop.rs
• crates/tui/src/seam_manager.rs
• crates/tui/src/models.rs
• docs/CONFIGURATION.md
• docs/capacity_controller.md
• related issues: Unified layered context management with Flash seam manager (v0.7.2) #159, v0.7.5: audit compaction and seam timing against V4 prefix cache #200, v0.7.5: simplify context management and make capacity docs match code #201, Recalibrate capacity controller entropy priors for V4 architecture #162

Questions to answer

• Are any context-management decisions still using lifetime summed usage where they should use active request size?
• Are UI/status/cost labels clear about the difference between current active context, cumulative billed prompt tokens, output tokens, and cache hit/miss tokens?
• Do hard cycle, optional seam production, emergency overflow recovery, manual compaction, and capacity-controller checkpoints all use the correct token basis for their job?
• Are tests explicit enough that future agents will not reintroduce cumulative-usage based triggers?

Acceptance checks

• Add/adjust tests that fail if hard-cycle or seam triggers use lifetime summed prompt usage.
• Add a small audit note or doc paragraph naming the token bases used by each subsystem.
• Preserve v0.7.5 policy: V4 default remains append-only/cache-preserving; do not enable replacement compaction, Flash seams, or capacity intervention by default.
• Do not close Unified layered context management with Flash seam manager (v0.7.2) #159, v0.7.5: audit compaction and seam timing against V4 prefix cache #200, v0.7.5: simplify context management and make capacity docs match code #201, or Recalibrate capacity controller entropy priors for V4 architecture #162 unless the implementation truly resolves their acceptance criteria.

Non-goals

• No provider expansion.
• No Whalescale rename/migration.
• No Cargo/npm publish.

[Hmbown]

Closing as implemented. Evidence: crates/tui/src/cycle_manager.rs and crates/tui/src/seam_manager.rs document/use active request input estimates rather than lifetime summed API usage; regression tests include should_not_count_lifetime_api_usage_as_active_context and seam_trigger_uses_active_request_size_not_lifetime_usage. Verification run: cargo test -p deepseek-tui --bin deepseek-tui context --locked passed.