fix(session): invalidate token usage after pruning tool output

[lewismosciski]

Change Summary

Tool-output pruning retains later assistant usage measured over the original output. The next strategy therefore still sees the old token count and can call the summary model even when pruning brought the context below its limit.

Discard those stale usage fields from the pruned view. Keep measurements preceding the first pruned output, and preserve the original session records.

Validation: three regressions failed before the fix; 153 related tests passed afterward, with 20 skips. The assembler regression verifies that no summary call occurs and the corrected view survives reopening the session. Applicable changed-file pre-commit hooks passed. Full-repository pre-commit still reports existing lint/format failures on main, including skills_manager.py:61 and shell_validator.py:91.

Related issue number

Found while inspecting the default context-compaction pipeline; no existing issue.

Checklist

• The pull request title is a good summary of the changes - it will be used in the changelog
• Unit tests for the changes exist
• Run pre-commit install and pre-commit run --all-files before git commit, and passed lint check. (Both run; existing repository failures noted above.)
• Documentation reflects the changes where applicable (internal fix; no API or configuration change)