fix(agent): invalidate flush-scan cursor when finalizer pops db marker

[spfcraze]

What does this PR do?

Fixes the delivered final response silently going missing from the durable transcript when a turn ends on a pure tool-call tail.

Two mechanisms collide:

1. finalize_turn's tail fill (agent/turn_finalizer.py): when the transcript tail is a pure tool-call assistant row with empty content but a final_response was delivered, it fills that row's content in place and pops _db_persisted — explicitly so the next _persist_session re-writes the filled content (the fix for Gateway can deliver assistant response without persisting assistant row NousResearch/hermes-agent#43849/Telegram: assistant responses not persisted to session DB (model re-answers old messages) NousResearch/hermes-agent#44100).
2. The bounded flush scan (run_agent.py:_flush_messages_to_session_db_unlocked, perf commit 30c783589): skips the identity-matched prefix of the previous flush's snapshot, on the documented assumption that "no code path pops _DB_PERSISTED_MARKER from a live dict in place."

That assumption is false — the tail fill is the one place that does. The filled row is the same dict object flushed mid-turn, so the identity match skips it and the marker pop has no effect: the user sees the answer, but state.db keeps content="" and /resume replays an unanswered backlog, so the model re-answers — the exact NousResearch#43849/NousResearch#44100 symptom, resurfacing via the perf cursor.

The fix invalidates the cursor at the pop site (agent._db_flush_scan_prefix = None), so the filled row is re-examined. The perf optimization is untouched everywhere else; the append-only flush design is preserved (the re-persist appends the filled row, same as pre-cursor behavior).

Related Issue

No GitHub issue — discovered via code review and reproduced live (see below). Happy to file one first if preferred.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• agent/turn_finalizer.py: after popping _db_persisted in the tail fill, invalidate agent._db_flush_scan_prefix (with a comment explaining the interaction).
• tests/agent/test_turn_finalizer_final_response_persistence.py: new test — the fill path invalidates the cursor (and pops the marker) on a pure tool-call tail.
• tests/run_agent/test_identity_flush.py: new end-to-end test — incremental flush stamps the tool-call row; after fill + pop + invalidation, the turn-end flush persists the filled answer to a real SessionDB.

How to Test

Reproduction (against pre-fix code, real AIAgent + real SessionDB):

agent._flush_messages_to_session_db_unlocked(messages)   # mid-loop: tool-call row written, stamped, cursor set
messages[0]["content"] = "the final answer"              # what finalize_turn's fill does
messages[0].pop("_db_persisted", None)
agent._flush_messages_to_session_db_unlocked(messages)   # turn-end persist
# pre-fix:  DB contents are ['', 'next question']  — the answer never lands
# post-fix: DB contents are ['', 'the final answer', 'next question']

Focused validation completed:

1. bash scripts/run_tests.sh tests/agent/test_turn_finalizer_final_response_persistence.py tests/run_agent/test_identity_flush.py — 9/9 pass.
2. Sabotage check: stashing the agent/turn_finalizer.py change makes exactly the cursor-invalidation test fail (8 pass); restoring returns to 9/9.
3. Wider suites: all tests/agent/test_turn_finalizer_* + the full tests/run_agent/ directory — 1278/1278 pass.
4. Full repo-wide suite, bash scripts/run_tests.sh (branch): 22,817 pass / 100 fail. Baseline on clean main (this branch's parent), same machine: 22,816 pass / 99 fail — failing sets identical except tests/honcho_plugin/test_pin_peer_name.py, which fails identically standalone on clean main (pre-existing flake, no agent-persistence imports). Nothing in the persistence/finalizer surface fails.
5. uvx --from ruff==0.15.10 ruff check agent/turn_finalizer.py tests/agent/test_turn_finalizer_final_response_persistence.py tests/run_agent/test_identity_flush.py — clean.
6. git diff --check — clean.
7. Adversarial cases executed: pre-cursor behavior parity (the re-persist appends the filled row, matching the append-only design); marker-pop without fill (no tail fill) is unaffected; the existing test_final_response_fills_pure_tool_call_tail still passes unchanged.

The full repo-wide suite was run locally (item 4) with results verified against clean main; GitHub CI remains the final confirmation environment.

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 (keyword matrix + same-file API sweep of open PRs touching run_agent.py / agent/turn_finalizer.py — closest were the in-flight-marker family fix(agent): preserve in-flight marker on flush failure NousResearch/hermes-agent#74305/fix(state): don't clear in-flight turn marker when canonical flush fails NousResearch/hermes-agent#74540 and fix(agent): do not re-persist empty failure sentinel after scaffolding drop NousResearch/hermes-agent#69279's sentinel drop, all different defects)
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run the full suite locally (scripts/run_tests.sh): 22,817 pass / 100 fail — failures verified identical to clean main (environment lanes; not all pass, so the stock claim above is restated accurately here — see How to Test, item 4)
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Ubuntu 24.04

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A (inline comment documents the cursor/marker interaction)
• 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 — no platform surface
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

Logs

Sabotage verification output:

# fix stashed (pre-fix behavior):
=== Summary: 2 files, 8 tests passed, 1 failed ===   (the cursor-invalidation test)
# fix restored:
=== Summary: 2 files, 9 tests passed, 0 failed ===