[Pepps233]

Summary

Reduce repeated goal-mode context usage notes by reporting after the first completed tool batch and then every other batch.

Reasoning

Routine usage metadata does not need to repeat after every batch.
Keep context markers and request-fit checks on every batch, and report higher pressure tiers immediately so compaction remains timely.
Derive cadence from persisted request boundaries so retries do not advance it or rewrite history.

Changes

• Alternate usage and guidance notes during tool loops; refresh for new input and request configuration changes.
• Record the last reported usage in boundary metadata and validate it on replay.
• Preserve per-exchange markers, compaction eligibility checks, and emergency recovery.
• Document cadence and cover normal/error batches, pressure thresholds, retries, and sustained-loop marker stability.

Testing

• npm run check.
• Focused faux-provider tests for cadence, markers, cache stability, steering, compaction, and recovery.
• ./test.sh: all scripts and other packages passed; found and updated one test's old every-batch expectation.
• ./test.sh coding-agent: 3,380 passed, 49 skipped after updating that expectation.