fix(agent): strip _db_persisted on rotation compression assembly (salvage #57531, #57491)

[kshitijk4poor]

Summary

Salvage of #57531 (@nankingjing) — cherry-picked to preserve authorship, rebased clean onto current main, and independently verified end-to-end. Fixes #57491.

Rotation compression (compression.in_place: false) on long-lived cached gateway sessions silently drops the compacted transcript from state.db. On restart the child session reloads a near-empty transcript → severe amnesia. Regression introduced by #50372 (e4c6d1b22), which switched flush dedup from id(msg) to an intrinsic _db_persisted marker.

Root cause: ContextCompressor.compress() assembles the post-rotation list with shallow messages[i].copy() for the protected head (context_compressor.py:2837) and tail (:2910). On a cached agent every live message already carries _db_persisted (stamped by incremental flushes), and .copy() propagates it. The child-session flush (run_agent.py:1820, if msg.get(_DB_PERSISTED_MARKER): continue) then skips every copied row — only brand-new post-compression appends reach the child DB.

Fix: strip _db_persisted when copying head/tail into the compressed transcript, via a small _fresh_compaction_message_copy() helper applied at both assembly sites. In-place compaction is untouched.

Independent verification (not just green units)

Reproduced against the real _flush_messages_to_session_db + real SessionDB on a temp HERMES_HOME, both variants:

Variant	Unfixed main	With fix
Standalone summary	child gets 2/9 rows	9/9 ✓
_merge_summary_into_tail (summary folded into a marked tail dict)	child gets 1/8 rows, 0 [CONTEXT COMPACTION] rows	8/8, summary present ✓

The merge-into-tail row exactly matches the WeChat production evidence in the issue ("Child #51: 17 messages, 0 rows containing [CONTEXT COMPACTION]") — the worst case where the summary itself is lost. The fix covers it because the tail copy is now stripped before the summary is merged into it.

Chokepoint check: the two post-assembly passes (_sanitize_tool_pairs, _strip_historical_media) operate on the already-stripped compressed list and never re-copy from the original messages, so sites 2837/2910 are the complete chokepoint — no marker can re-enter.

Mutation-tested: neutering the strip (revert helper to plain .copy()) makes test_rotation_child_session_flushes_full_compressed_transcript_with_markers fail — the regression test genuinely guards the fix.

Note on the local constant: the helper redefines _DB_PERSISTED_MARKER locally rather than importing it from run_agent — importing back would be circular (run_agent imports context_compressor at import time). Consolidating both literals into a shared constants module is a larger refactor out of scope for this bug.

Tests

• tests/agent/test_context_compressor.py — unit: assembled compressed list carries no _db_persisted
• tests/run_agent/test_compression_persistence.py — e2e: cached agent with all source messages pre-marked → rotation → child receives the full compressed row count

$ pytest tests/agent/test_context_compressor.py tests/run_agent/test_compression_persistence.py \
         tests/agent/test_context_compressor_session_end_clears_state.py \
         tests/agent/test_context_compressor_summary_continuity.py \
         tests/agent/test_context_compressor_cross_session_guard.py -q
164 passed

Supersedes #57531 and #57508 (both independent fixes for the same root cause; credit to @nankingjing and @rayjun).

[alt-glitch]

This was generated by AI during triage.

Duplicate of #57508 (canonical, earliest of the same-mechanism cluster). This PR is a salvage of #57531 and fixes the same issue #57491 via the same mechanism (strip the _db_persisted marker from messages[i].copy() in context_compressor.py compressed-transcript assembly). Cluster fixing #57491: #57508 (earliest, 03:57), #57531 (salvage source, 04:49), #57574 (this, salvage). Maintainer to pick canonical; related family salvage: #48584.