fix(memory): preserve existing DB when sqlite3_open fails

[cursor]

Summary

• memory_init no longer remove()s an existing SQLite file when sqlite3_open fails. Permissions or transient I/O no longer destroy sessions, long-term memories, or cron jobs.
• Reimplemented on current development (includes fix(providers): fail closed when Anthropic parse realloc fails #80). Not a cherry-pick of the original Bot commit.
• Isolated to src/core/memory.c and tests/test_memory.c. Missing-file open failures still unlink and retry.

Test plan

• test_existing_db_preserved_on_open_failure (red: chmod 000 open-fail path deleted and recreated the DB; green: init returns -1 and data is intact)
• make test_memory
• CI=true make test
• make static

Refs: #63

---

Note

Medium Risk
Changes startup behavior for the persistent memory store; failures now fail closed without wiping data, which is safer but leaves the agent unable to start until the DB is readable again.

Overview
memory_init no longer deletes an existing SQLite file when sqlite3_open fails (e.g. permissions or transient I/O). It logs the error, closes the handle, and returns -1 instead of remove() + recreate. The unlink-and-retry path is unchanged when the DB path did not exist before open.

A new unit test (test_existing_db_preserved_on_open_failure) chmods the DB to block opens, asserts init fails without data loss, then restores permissions and verifies recall still works. CHANGELOG documents the fix.

^{Reviewed by Cursor Bugbot for commit 5f8d420. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Stale comment

Could not approve: GitHub rejects a self-approval because this bot authored the PR. Cursor Bugbot and Cursor Security Agent both passed with no findings that need human review. adriannoes is already assigned; no additional reviewers were added.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}

[cursor]

Summary

ShellClaw is a C99 edge agent (Jetson / Raspberry Pi) with sessions, long-term memories, and cron in one SQLite file. This PR closes the remaining data-loss hole in memory_init: commit 4268d69 already refused to delete an existing DB on schema failure (CR-18), but sqlite3_open failure still called remove() and retried. Five lines in src/core/memory.c now return -1 when path_exists was true, matching that schema-mismatch path. init_subsystems already aborts startup on that -1. I ran test_memory as uid 1000; all tests passed, including the new case.

Must Fix

None. Production behavior matches the PR claim. Missing-path open failures still unlink and retry. No injection, secrets, or new coupling.

Should Fix

1. src/core/memory.c (new fprintf on the open-fail path): capture sqlite3_errmsg(g_db) before sqlite3_close. After close the message is gone. Operators on EACCES or I/O currently get only the path, then bootstrap's generic memory init failed.
2. tests/test_memory.c test_existing_db_preserved_on_open_failure: chmod(path, 0000) does not block sqlite3_open as root. sudo ./build/test_memory fails at ASSERT(memory_init(path) == -1). GitHub ubuntu-24.04 is non-root, so required CI should pass. Skip when geteuid()==0, or occupy the path with a directory (sqlite3_open fails even as root; the old remove() of an empty dir would still recreate). Restore 0600 if an ASSERT fires after chmod; a failed root run left a mode-000 file in /tmp.

Nice to Have

• Document fail-closed existing files in memory.h (memory_init still only says it creates the file if missing).
• A SHELLCLAW_TEST open-fail hook would be root-proof, same idea as the Anthropic realloc hook in #80.
• test_corrupted_db_recreated still has a name that implies recreate. Pre-existing; garbage files usually fail in run_schema, not sqlite3_open.

Positive Highlights

• Same file_existed gate as the schema path. No new control-flow style.
• The regression test would have been red on the old unlink-and-recreate: the owner can unlink a mode-000 file they own, then sqlite3_open creates an empty DB and init returns 0 with data gone.
• Isolated to memory.c + test_memory.c, CHANGELOG Unreleased Fixed line in a separate docs commit, Conventional Commits, no secrets.

  

_{Sent by Cursor Automation: Adrianno’s personal code review}

[cursor]

Could not approve: GitHub rejects a self-approval because this bot authored the PR. Cursor Bugbot and Cursor Security Agent both passed with no findings that need human review. adriannoes is already assigned; no additional reviewers were added.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}