fix(session): avoid O(n²) semantic re-processing on memory commit

[sliverp]

Problem

When a new memory is extracted via session.commit(), the session unconditionally enqueues a SemanticMsg without any changes information (line 315 of session.py). This causes _process_memory_directory() to re-summarise and re-vectorise every file in the memory directory, regardless of whether it changed.

The cumulative cost grows as O(n²) with memory count:

• Store 1st memory → process 1 file
• Store 100th memory → process 100 files
• Store 500th memory → process 500 files
• Total = 1 + 2 + ... + n = O(n²)

Real-world impact: ~500 memories consumed hundreds of thousands of embedding tokens per day on Volcengine, with 2000+ rate-limit (429) retries.

Root Cause

Two interacting issues:

1. session.py:315 — commit_async() always enqueues a full-directory SemanticMsg even when SessionCompressor._flush_semantic_operations() has already enqueued incremental messages with per-file change sets.

2. semantic_processor.py:369 — _process_memory_directory() only attempts to reuse existing summaries from .overview.md when msg.changes is not None. When the fallback path sends a message without changes, every file gets re-summarised.

Fix

1. session.py: Only enqueue the fallback SemanticMsg when the compressor is absent or extracted 0 memories. When the compressor runs successfully, it already handles incremental semantic processing.

2. semantic_processor.py: Always try to load existing summaries from .overview.md, regardless of whether msg.changes is set. This ensures even the fallback/redo path can skip unchanged files.

Tests

Added 4 unit tests in tests/unit/session/test_incremental_semantic.py:

• test_commit_skips_fallback_semantic_when_compressor_flushed — verifies no duplicate SemanticMsg
• test_commit_enqueues_fallback_semantic_when_no_compressor — verifies fallback still works
• test_commit_enqueues_fallback_when_compressor_extracts_zero — edge case
• test_semantic_msg_changes_none_by_default — documents the default

Impact

For a workspace with N memories, this reduces per-commit semantic processing from O(N) file summaries down to O(changed_files), and eliminates the O(N²) cumulative cost growth.

Closes #505
Ref #744

[CLAassistant]

All committers have signed the CLA.

[myysy]

Thanks for the PR. The final semantic step in commit_async targets the session path, which is different from the memory path flushed by the compressor, so it can’t be removed or skipped. We’ll revisit and adjust the session-path semantic logic later.

[github-actions]

Failed to generate code suggestions for PR