fix(agent): stop re-appending repaired rows and saving the length nudge as user text

[jonpol01]

What does this PR do?

A repaired user row is saved once, and Hermes's internal length-continuation nudge is no longer saved as the user's own words.

Before this change, #117569 popped a saved row's marker so that "the next flush rewrites it". But the flush only appends, so the user's prompt was inserted a second time and every later turn replayed both. The thinking-only length retry and the interrupted lease-wait carry both hit it.

Related Issue

Fixes #120861

This fixes forward from #117569 (#117120, salvage of #117121) and keeps its intent: an in-place repair of a persisted dict reaches state.db, now as an in-place rewrite. Nothing is reverted.

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

• A. agent/transcript_repair.py: a message that already owns an active row (_row_id) is never inserted again.

  • A non-assistant row is rewritten in place.
  • Assistant rows keep the existing blank-update / concurrent-winner logic.
  • _active_assistant_row becomes _active_row(role), with the same watermark-clone lookup.

• B. agent/turn_truncation.py, session_persistence.py, turn_final_response.py: a length nudge that would follow this turn's already-saved user row is flagged _length_continuation_synthetic. That makes it request-only ephemeral scaffolding, like _dropped_toolcall_nudge: the wire copy still merges it for alternation, but it is never saved or merged into the real row, and it's removed before the final answer.

  • Skipping the nudge outright would break test_empty_stub_only_appends_continuation_user_message, which requires the chunking nudge to reach the model.

• C. agent/session_persistence.py: _PERSIST_AFTER_ADMISSION_INTERRUPT is cleared after its one write, so the carried row stops bypassing the history skip.

• agent/agent_runtime_helpers.py _merge_consecutive_users never merges ephemeral scaffolding into a real user row.

• Tests (tests/agent/test_session_persistence_repaired_user_row.py) drive the real turn loop with a fake LLM client and a real SessionDB. No request leaves the machine.

  • test_thinking_only_length_retry_saves_the_prompt_once_and_never_the_nudge
  • test_prompt_carried_past_an_interrupted_lease_wait_is_saved_once: a real held lease, the status callback, and interrupt().
  • Both are red on main, with the saved prompt duplicated and the nudge saved as user text. Reverting each piece of B and C on its own reddens a test.
  • Piece A was checked outside these two tests with two real-loop runs:
    • a tool result with a lone surrogate, followed by UnicodeEncodeError recovery, saved one tool row instead of two;
    • In-place mutations of _db_persisted live dicts leave stale rows in state.db #117120's own merged-row case gives one row with the merged text, not 'first' plus 'first\n\nsecond'.

  Happy to add either as a third test if wanted.

Trade-off: after such a continuation, the next turn replays the prompt without the nudge, so the prompt cache misses from that message onward, exactly as _dropped_toolcall_nudge already does. Before, the corrupted merged text was what got cached.

Open PRs:

• fix(agent): recover a router-hidden truncated tool call instead of ending the turn partial #111638 adds a chunking nudge it expects to be folded in; it should set _length_continuation_synthetic when it follows a user row.
• fix(agent): pop persisted marker on repair mutations #104454 re-adds the same pop next to the new condition, so expect a conflict.
• fix(session-db): persist post-flush /steer mutation to SQLite #52604 builds its own row-update mechanism that overlaps piece A.

How to Test

1. scripts/run_tests.sh over test_session_persistence_repaired_user_row.py and the fix(agent): in-place repairs of persisted message dicts reach state.db instead of leaving stale rows (#117120, salvage #117121) #117569 suites (test_message_sequence_repair, test_micro_compaction, test_tool_call_args_sanitizer, test_unicode_ascii_codec), plus test_partial_stream_finish_reason: 143 passed.
2. 78 related files over continuation, lease, batch, scaffolding, repair and _row_id: 1918 passed.
   • The two failures in test_run_agent.py and one in tui_gateway fail identically on unmodified main on this host: the optional anthropic SDK is missing, and a git-probe fork segfaults.
3. Hunt repro: the saved transcript after turn 2 is [user 'first question', assistant, user 'write me a long report', assistant], with the prompt once and no nudge row.

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 ran only the related files above through scripts/run_tests.sh, not the full suite.
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 26

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