fix(cli): erase clears all builder-owned files (#349, #350)

[HumanBean17]

What

Follow-up to #346 / PR #348. erase removed code_graph.lbug, cocoindex.db, and .graph_hashes.json but left the rest of the graph builder's bookkeeping on disk:

• Bug: erase leaves .graph_increment_in_progress; next increment silently does a full rebuild #349 — .graph_increment_in_progress (the incremental crash marker) survived erase and init. The next increment then saw the stale marker and silently fell back to a full rebuild — a one-time perf hit explained only under --verbose.
• Bug: erase leaves .graph_hashes.json.tmp orphan after a crashed build #350 — .graph_hashes.json.tmp (the atomic-write temp from FileHashTracker.save) was orphaned after a crashed save. Pure cruft, but it defeated the "clean slate" erase advertises and never appeared in the Will delete: preview.

Fix

_cmd_erase hardcoded only .graph_hashes.json. Per both issues' suggestion, export the builder-owned filenames as a single source of truth:

• build_ast_graph.BUILDER_OWNED_INDEX_FILES = (GRAPH_HASHES_FILENAME, GRAPH_HASHES_TMP_FILENAME, GRAPH_INCREMENT_MARKER_FILENAME) — defined next to the FileHashTracker / crash-marker code that owns them.
• FileHashTracker and incremental_rebuild now reference the constants instead of literals.
• _cmd_erase derives its delete + preview list from BUILDER_OWNED_INDEX_FILES, so erase and the builder cannot drift.

Test

New test_erase_removes_increment_marker_and_hash_store_tmp seeds the crash marker + the orphan temp and asserts erase removes both. TDD: failed first (red on the marker assertion), then green after the fix.

.venv/bin/python -m pytest tests/test_java_codebase_rag_cli.py -q
# 62 passed

Notes

• No re-index required, no env-var change, no ontology_version bump — pure lifecycle/CLI cleanup.
• build_ast_graph was previously only run as a subprocess; this is the first import build_ast_graph. Verified import-safe: main() is __main__-guarded and there are no top-level side effects.

Closes #349, closes #350.

🤖 Generated with Claude Code

[HumanBean17]

Addressed a finding from a multi-agent review pass:

Top-level build_ast_graph import cost (low) — from build_ast_graph import BUILDER_OWNED_INDEX_FILES at module top pulled numpy/ladybug/pyarrow/tree_sitter on every CLI invocation (~54ms; -X importtime: cli cumulative 44ms → 98ms), including --help, violating cli.py's own lazy-import header. Moved the import into _cmd_erase (the only user). After the move build_ast_graph is not imported at cli import time and cumulative is back to ~39ms. test_java_codebase_rag_cli.py 62 passed.