Summary

Improve token-estimate accuracy so the /context view and compaction triggers agree with reality. The current estimator is a len/4 chars-per-token heuristic (+5 overhead) in compaction.EstimateMessageTokens.

Proposed behavior

• Reconcile the heuristic against provider-reported usage per message (already available via MessageUsage, pkg/runtime/event.go ~line 372).
• Keep a conservative bias so we compact slightly early rather than overflow.

Implementation notes (starting points)

• pkg/compaction/compaction.go
• pkg/runtime/compactor/compactor.go
• Extend estimator tests

Complexity: S–M.

Background: shared plan context-management-gap-analysis (Phase 2, item 6).

Dependencies / Relationships

• Related: feat(tui): add /context command with a categorized context-window breakdown #3432 (improves accuracy of the breakdown numbers), feat(config): make the auto-compaction threshold configurable (compaction_threshold) #3433 (compaction-trigger accuracy)