fix(embeddings): enumerate nodes independently of stored paths

[y4ho0]

Linked issue

Closes #945

What & why

Graphs with legacy Windows path spellings can produce zero embeddings because node enumeration passes stored paths through normalized file lookups. The tool then incorrectly reports that semantic search is active.

Explicit embedding and opted-in refresh now enumerate all non-File nodes directly, including virtual nodes outside the file inventory. The summary reports that semantic search is unavailable when the resulting embedding count is zero; an already-current index still reports active. Stored paths, node identities, and provider/model checks are preserved.

How it was tested

Python 3.13.15 on macOS. Regression tests use real SQLite stores with simulated legacy Windows paths and deterministic embedding providers. They cover initial embedding, explicit refresh, repeat runs, virtual nodes, and empty/File-only graphs.

uv run --frozen pytest tests/ --tb=short -q --disable-warnings
# 2993 passed, 9 skipped, 2 xpassed.
uv run --frozen ruff check code_review_graph/ tests/test_embedding_refresh.py
# All checks passed.
uv run --frozen mypy code_review_graph/ --ignore-missing-imports --no-strict-optional
# Success: no issues found in 72 source files.

The coverage run reached 85.06% (65% required), with 2,992 passed, 9 skipped, 2 xpassed, and one failure in the unchanged real filesystem-watcher test test_recreated_directory_survives_and_is_reindexed. That test passed in isolated reruns on both this branch and clean upstream. An initial Python 3.14 run also hit the existing Windows event-loop test failure, reproduced on the unmodified base; final validation uses CI-supported Python 3.13.

Checklist

• Regression tests added
• Full test suite passes on Python 3.13
• Linting passes
• Type checking passes
• Changed lines are at most 100 characters
• Tool return documentation describes the zero-embedding summary

[github-actions]

code-review-graph review

Overall risk: 0.40 (MEDIUM) — 8 changed function(s)/class(es), 17 affected flow(s), 1 test gap(s)

Risk-scored changes

Risk	Level	Symbol	Location	Tested
0.40	medium	code_review_graph/tools/docs.py::embed_graph	code_review_graph/tools/docs.py:28	yes
0.30	low	code_review_graph/embeddings.py::embed_all_nodes	code_review_graph/embeddings.py:1307	yes
0.30	low	code_review_graph/embeddings.py::refresh_embeddings	code_review_graph/embeddings.py:1318	yes
0.05	low	tests/test_embedding_refresh.py::TestEmbeddingNodeEnumeration	tests/test_embedding_refresh.py:114	no
0.05	low	tests/test_embedding_refresh.py::TestEmbeddingNodeEnumeration.test_embeds_non_file_nodes_outside_the_file_i...	tests/test_embedding_refresh.py:115	(test)
0.05	low	tests/test_embedding_refresh.py::TestEmbeddingNodeEnumeration.test_manual_embed_reaches_nodes_regardless_of_...	tests/test_embedding_refresh.py:134	(test)
0.05	low	tests/test_embedding_refresh.py::TestEmbeddingNodeEnumeration.test_refresh_reaches_changed_nodes_regardless_o...	tests/test_embedding_refresh.py:152	(test)
0.05	low	tests/test_embedding_refresh.py::TestEmbeddingNodeEnumeration.test_tool_summary_reflects_whether_any_vectors...	tests/test_embedding_refresh.py:173	(test)

Affected execution flows

• embed_query — criticality 0.48, 2 node(s) across 1 file(s)
• embed_query — criticality 0.48, 2 node(s) across 1 file(s)
• embed_query — criticality 0.45, 3 node(s) across 1 file(s)
• embed_query — criticality 0.45, 3 node(s) across 1 file(s)
• embed_query — criticality 0.41, 2 node(s) across 1 file(s)
• ...and 12 more affected flow(s)

Test gaps

• tests/test_embedding_refresh.py::TestEmbeddingNodeEnumeration (tests/test_embedding_refresh.py:114)

Token savings: this graph-backed report used ~14,880 fewer tokens (~69%) than reading every changed file in full (estimated, chars/4 approximation).

---

Powered by code-review-graph — local-first analysis; no code leaves the CI runner.

[tirth8205]

Integrated on integration/token-efficiency-hardening as f68182d. Embedding refresh now enumerates non-File nodes directly, retaining legacy paths and virtual nodes, and reports zero-vector availability accurately. Targeted refresh tests passed with the existing provider opt-in rules preserved; two unchanged loopback-server tests were unavailable in the sandbox.