perf: force condense ended orphaned ended sessions

[peyton-alt]

Fixes O(N) commit-time overhead caused by stale ENDED sessions accumulating and being re-processed on every
PostCommit (~73-103ms per session, indefinitely). Tracked as issue #591.

Two complementary fixes:

• Eager condense on session stop (CondenseAndMarkFullyCondensed): When a session ends via the stop hook and
  has no FilesTouched, its checkpoint data is immediately condensed to entire/checkpoints/v1 and the session
  is marked FullyCondensed. PostCommit skips these sessions entirely on all future commits.
• Force-condense safety net in PostCommit: ENDED sessions that are older than 1 hour and still have
  FilesTouched with no commit overlap are now force-condensed instead of being carried forward forever. This
  catches sessions the eager path couldn't handle (older CLI versions, crashes, or sessions that had
  FilesTouched at stop time — the primary subagent accumulation shape from Orphaned sessions cause 3GB+ RAM usage and 2m44s commit times #591).

Key design choices:

• CondenseAndMarkFullyCondensed intentionally keeps Phase = ENDED (unlike CondenseSessionByID which resets
  to IDLE) and skips sessions with FilesTouched so PostCommit's carry-forward tracking still works for those
• The 1-hour forceCondenseThreshold gives users a window to commit a session's files in a separate commit
  before the safety net fires
• Force-condensed sessions skip carry-forward computation entirely — FilesTouched is cleared to nil

Test plan

• mise run test — unit tests including new
  TestPostCommit_EndedSessionCarryForward_ForceCondensedWithoutOverlap,
  TestPostCommit_RecentEndedSession_NotForceCondensed, TestPostCommit_EmptyEndedSession_MarkedFullyCondensed,
  TestCondenseAndMarkFullyCondensed_Guards, TestCondenseAndMarkFullyCondensed_WithDataNoFiles
• mise run test:integration — TestCarryForward_NewSessionCommitDoesNotCondenseOldSession,
  TestSubagentAccumulation_Issue591
• mise run test:ci — full suite including E2E canary
• Verify accumulation_bench_test.go regression: second PostCommit is >2x faster than first once sessions
  are FullyCondensed

[cursor]

PR Summary

^{Cursor Bugbot is generating a summary for commit 091a9a4. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Comment @cursor review or bugbot run to trigger another review on this PR

[Copilot]

Pull request overview

This PR addresses a PostCommit performance issue where ENDED sessions with carry-forward FilesTouched but no overlap with later commits could be re-processed indefinitely, causing cumulative per-commit overhead (issue #591). It introduces a “force-condense” path to ensure those sessions get condensed once and then skipped on subsequent commits.

Changes:

• Add a force-condensation path for ENDED sessions with FilesTouched and new content even when there’s no commit overlap.
• Update PostCommit phase tests to assert the new force-condense behavior (cleared FilesTouched, reset counters, FullyCondensed set).
• Add benchmarks to measure PostCommit overhead with accumulated ENDED sessions and verify the “second commit after fix” performance.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
cmd/entire/cli/strategy/manual_commit_hooks.go	Adds forceCondensed flow to condense ENDED sessions without overlap and skip carry-forward.
cmd/entire/cli/strategy/phase_postcommit_test.go	Updates test expectations to confirm ENDED sessions are force-condensed and marked FullyCondensed.
cmd/entire/cli/strategy/accumulation_bench_test.go	Adds benchmarks and repo/session setup helpers to quantify accumulation overhead and validate the fix.