fix(acp): preserve compaction-archived transcript on per-turn persist

[mrparker0980]

What does this PR do?

Stops every long-running ACP session from silently losing its entire
pre-compaction transcript. The ACP AIAgent is created with the same
session_id and SessionDB as the adapter (_make_agent), so once its
context grows past the compression threshold it compacts in place via
SessionDB.archive_and_compact(), which deliberately soft-archives the
pre-compaction turns (active=0, compacted=1) to keep them on disk and
discoverable by session_search (the #38763 durability guarantee). The
problem: as soon as that turn ends, SessionManager._persist() called
replace_messages(), whose unconditional DELETE FROM messages WHERE session_id = ? wiped all rows for the id — including the ones the agent
had just archived. Net effect: auto-compaction fires, the turn returns, and
the archived history is gone.

Why this approach: the per-turn persist only ever needs to refresh the live
(active) set — the agent already owns archival. So _persist now replaces
only active=1 rows and leaves soft-archived rows untouched, rather than
trying to teach the adapter about compaction state. This keeps the existing
atomic delete+reinsert contract (the #13675 rollback-on-failure guarantee)
fully intact for the live set.

Why not change replace_messages outright: it's also used by destructive
rewrite flows (/retry, /undo, /compress, forks) that genuinely want every
row gone, so the new behavior is opt-in via an active_only flag and all
other callers are unchanged.

Related Issue

N/A

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• hermes_state.py: add an active_only: bool = False parameter to
  SessionDB.replace_messages(). When set, the DELETE is scoped with
  AND active = 1 so soft-archived rows survive; message_count/
  tool_call_count then track the live set, matching archive_and_compact().
  Default is unchanged, so existing callers keep their destructive semantics.
• acp_adapter/session.py: SessionManager._persist() now calls
  replace_messages(..., active_only=True), and the method docstrings are
  updated to reflect that only the live rows are replaced.
• tests/acp/test_session.py: add
  test_save_session_preserves_compaction_archived_history, a regression
  test that compacts a session in place and asserts the archived turns remain
  retrievable via get_messages(include_inactive=True) after a subsequent
  save_session().

How to Test

1. Reproduce: with the fix reverted, run the new test — it fails because the
   per-turn persist deletes the archived rows.
2. Verify the fix: scripts/run_tests.sh tests/acp/test_session.py — the new
   test_save_session_preserves_compaction_archived_history passes, and the
   existing fix(acp): preserve persisted session history on save failures #13675 rollback test still passes.
3. Regression-check the shared method: the replace_messages /
   archive_and_compact SessionDB tests in tests/hermes_state/ and
   tests/test_hermes_state.py all pass, confirming default callers are
   unaffected.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 15 (Darwin 25.5)

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

[kshitijk4poor]

Closing as a duplicate of #50405, merged via #56342. Both PRs fixed the same ACP archived-history data loss. #50405's owned-agent skip also handles the compression id-rotation case; your active_only=True primitive was adopted (with credit) in the follow-up commit 723ccda to cover the model-switch / restore save paths. Thanks @mrparker0980 — your approach directly shaped the final fix.