fix(symbols): run symbol work in a spawned worker process - #54

Problem

Symbol builds and queries are CPU-bound Python. In the shared daemon they hold the GIL, so every other tool call waits behind a build. After #53, search_text against Vision still took 17.0 s during a 17.1 s symbol build, against 0.47 s alone.

Change

All symbol work runs in one persistent worker process started with spawn (src/lgrep/symbol_worker.py):

• index checks, refreshes, and bootstraps; explicit index_symbols_folder / index_symbols_repo jobs;
• search_symbols, get_symbol, get_symbols, search_references, get_file_outline, get_repo_outline;
• list_repos, invalidate_cache, and prune_symbols, which parse index bodies or evict the worker's index cache.

The daemon routes calls on the existing symbol lane and receives small picklable results. The worker:

• points stdout at stderr before loading any lgrep module that logs, and forwards its log events to the daemon's sinks with a worker_pid field;
• caps malloc arenas and trims its heap when idle;
• receives the shutdown stop event at start. A worker still running 3 s after shutdown begins is killed.

A crashed worker is replaced on the next call. Its in-flight calls return code="symbol_worker_restarted", retryable=true. Unchanged: single-flight refresh per path, a cancelled query does not stop its build, explicit-job status, counters and credential redaction, and the in-process CLI/Python path (InProcessSymbolWorker). No dependency or storage-format change.

The native memory hooks move from lgrep.server.memory to lgrep.memory, so the worker can use them without importing the server package.

Verification

• tests/test_symbol_worker.py (11 tests) covers search during a real spawned-worker build, an empty stdout, crash replacement, the stop event, the shutdown kill, and retryable tool errors. The responsiveness test fails when the build runs in the daemon: its one overlapping search took 2.45 s, the length of the whole build.
• Full suite: 1280 passed, 1 xfailed.
• Live stdio MCP measurement: search_text against Vision during a full index_symbols_folder build (164 files, 88,218 occurrences), fresh symbol store:

Cost

The worker is a second process: an idle worker holds about 161 MB RSS. Importing lgrep.storage loads LanceDB through the package __init__, which adds up to about 2 s to a cold worker start and part of that RSS. A follow-up covers it.