Problem

A consumer on Intel Mac (darwin + x86_64) reports jrag watch crashes: it tries to import sentence_transformers, which PEP 508 markers in pyproject.toml exclude on that platform (sentence-transformers / lancedb / cocoindex are gated to sys_platform != 'darwin' or platform_machine != 'x86_64'). The daemon prints failed to load embedding model: No module named 'sentence_transformers' and exits 2 before serving.

jrag watch was the only code path that hard-required the vector stack. Every other path already degrades to lexical/graph-only search (mcp_v2.search_v2 → _ensure_vector_backend() → ImportError → run_search=None → run_lexical_search; #443 made BM25/lexical first-class).

Root cause (two spots)

1. Startup — watch/daemon.py eagerly called self.warm.model() to fail-fast on model-load errors. On Intel Mac that pulls sentence_transformers → ModuleNotFoundError → exit 2. (The daemon's read path doesn't even use warm.model() — server.dispatch passes warm.graph() only; search_v2 loads the model lazily and falls back to lexical on its own. The eager call was purely fail-fast / first-query optimization.)
2. Reindex — watch/watcher.py:reindex always ran run_cocoindex_update and treated its nonzero rc as fatal. On Intel Mac that returns a 127 stub (cocoindex binary absent — no raise), so every debounced reindex emitted a perpetual "vectors" error and never fired indexing_done.

Fix

Reuse the existing pipeline.vector_stack_installed() probe (cheap: 3× importlib.util.find_spec) to gate both spots:

• watch/daemon.py — probe _vector_enabled in __init__; skip the model warm-up when vectors are absent (print a one-line lexical notice, continue serving). Add mode ("lexical" / "vector") to the state file and a panel row.
• watch/watcher.py — probe _vector_enabled in __init__; in reindex, skip the cocoindex step (and its fatal check) when vectors are off, so the graph reindex still completes and fires indexing_done.
• jrag.py — watch --status prints mode: lexical (graph-only) when applicable.

Outcome: the daemon now starts and stays useful on Intel Mac — serving warm lexical search (BM25 over the symbol graph) plus every structural command (find/inspect/callers/callees/flow), and reindexing the graph on file change.

Tests

• tests/watch/test_daemon.py — graph-only startup: a subprocess runs the real run_foreground with daemon.vector_stack_installed → False and a WarmResources.model() that raises if called (proving the gating); asserts the daemon reaches serving (not exit 2) and the state file carries mode: lexical.
• tests/watch/test_watcher.py — graph-only reindex({"java"}) skips cocoindex, runs the graph under its COW snapshot lifecycle, and fires indexing_done; a sql change is a clean no-op.
• Existing watcher tests pinned to _vector_enabled = True so the vector-path assertions are host-independent (don't implicitly depend on the host's vector stack).

Verification

• Watch + watcher tests: 36 passed, 6 skipped (heavy).
• Subset (watch/mcp/search/read_payloads): 413 passed, 26 skipped.
• Full suite: 1538 passed, 38 skipped (~8 min).
• Grep guard: no eager sentence_transformers/torch import anywhere in src/java_codebase_rag/watch/.
• End-to-end graph-only simulation (throwaway, on Apple Silicon): built a graph-only index → spawned the daemon with the vector stack forced off → jrag search returned hot-served (cold-loads=0), mode=lexical, rc=0, with real symbol results — impossible on the vector path since no vectors were built, proving the lexical path ran over the socket.

Docs

docs/JAVA-CODEBASE-RAG-CLI.md, docs/CONFIGURATION.md, docs/ARCHITECTURE.md note the lexical/graph-only mode. Skills/agents watch tips stayed accurate (warm lexical is still warm).

Out of scope

• No change to WarmResources.model() contract (only the daemon warm-up call is gated).
• No new CLI flag — capability is environment-determined, matching the cold CLI.
• pyproject.toml markers unchanged (Intel Mac stays graph-only by design).
• No version bump (separate concern in this repo).

Related: #435 (lexical-mode direction), #443 (lexical first-class).

🤖 Generated with Claude Code