fix: surface global config writes and correct first-run guidance

[karngyan]

Claim

Commands are honest about what they touch and guide the new user correctly. Four coherent UX fixes across setup, status, sessions, and search — output text only, no behavior or storage changes.

What changed

1. setup surfaces global config writes. When settings would be written to the global config file (AUTOPHAGY_CONFIG_DIR unset) while an explicit --database points elsewhere, setup now prints a notice before writing. Interactive runs fold it into the summary; --yes runs send it to stderr so JSON stdout stays clean. Config location is unchanged — only made visible. The "Re-running setup" line now keys off an on-disk config file rather than a possibly-stale flag.
2. status reports the real database size. The store opens in WAL mode, so a naive read of the main file right after migrations undercounts a brand-new database (~4 KiB header, not the ~250 KiB migrated schema). status now sums the -wal/-shm sidecars, and labels a file it just created as created new database.
3. Fresh-DB search hint. An empty database's search line points at autophagy setup (nothing to reindex yet). The reindex --index-tool-input hint stays for the imported-but-unindexed case where it is correct.
4. Friendly empty states. Empty sessions prints no sessions imported yet — run autophagy setup (no bare header); search against a zero-event database says so. --output json shapes are unchanged — empty arrays stay empty arrays.

Evidence

New CLI integration tests in crates/autophagy-cli/tests/cli.rs:

• status_reports_post_migration_size_for_a_new_database — size_bytes > 100_000 (was ~4 KiB), exists=false, text says created new database.
• status_search_hint_points_at_setup_before_import_and_reindex_after — empty DB → autophagy setup and no reindex; imported-unindexed → reindex --index-tool-input.
• empty_sessions_and_search_guide_the_user_but_keep_json_arrays — friendly text for both; JSON result stays []; a real miss on a populated DB still says no retrieval matches.
• setup_warns_before_writing_global_config_for_explicit_database — notice fires; the faked global path resolves under a throwaway HOME, never the real one.
• setup_omits_global_notice_when_config_dir_is_isolated — notice silent when AUTOPHAGY_CONFIG_DIR isolates the run.

Before / after (fresh empty database):

# status size — before: "4.0 KiB · schema v2 · new database (nothing imported yet)"
#                after:  "313.6 KiB · schema v2 · created new database (nothing imported yet)"

# status search line — before: "commands not indexed — run `autophagy reindex --index-tool-input`"
#                      after:  "commands not indexed — run `autophagy setup` to import your sessions"

# sessions — before: "SESSION\tSOURCE\tEVENTS\tLAST EVENT\tPROJECT"  (bare header)
#            after:  "no sessions imported yet — run `autophagy setup`"

# search   — before: "no retrieval matches"
#            after:  "no events imported yet — run `autophagy setup` to import your sessions"

# setup --database /tmp/x.db --yes (global config) — new stderr line:
#   note: settings are saved globally to <path> (set AUTOPHAGY_CONFIG_DIR to keep this run isolated)

Privacy implications

None. Output text only; no new persistence, network, or execution. Config write location is unchanged (now merely announced). The global-notice test fakes the global path via a throwaway HOME/XDG_DATA_HOME and never touches the real config or database.

Quality gate

mise run check green (ci + fmt + lint + test + docs).

🤖 Generated with Claude Code

https://claude.ai/code/session_015LY7xtercMc3NtLeX2z4YV