In `@tests/integration/hooks/helpers.ts`:
- Around line 8-9: Consolidate the two separate imports from 'child_process'
into a single import statement: replace the two lines importing spawnSync and
execFileSync separately with one combined import that imports both spawnSync and
execFileSync from 'child_process' (target the existing import declarations for
spawnSync and execFileSync in tests/integration/hooks/helpers.ts).
- Around line 96-97: The current shallow merge at the config creation line
(using defaults, overrides, and hooks: { ...defaults.hooks, ...overrides })
replaces nested hook objects (e.g., sessionStart) and loses inner properties
like memoryLimit; change this to a deep merge for hooks so nested properties are
preserved — either call a deep-merge utility (e.g., lodash.merge) or add a small
recursive merge helper and use hooks: mergeDeep(defaults.hooks, overrides),
keeping the top-level defaults merging logic intact (referencing the variables
config, defaults, overrides and the hooks/sessionStart objects).

In `@tests/integration/hooks/hook-init-hooks.test.ts`:
- Around line 37-57: This test depends on settings.json created by a previous
spec and can flake if run alone; modify the test in hook-init-hooks.test.ts to
ensure idempotency by creating or initializing the settings before asserting:
check for settingsPath (join(workDir, '.claude', 'settings.json')) and if
missing call the same setup helper or CLI init used by the earlier test (the
helper or command that creates .claude/settings.json—use the existing setup
function or test titled "should create .claude/settings.json and .git-mem.json"
or its underlying initializer), then read and assert the hooks as currently
written so the test can run independently.

In `@tests/integration/hooks/hook-prompt-submit.test.ts`:
- Around line 44-53: The test manually writes a full .git-mem.json instead of
using the writeGitMemConfig helper; update
tests/integration/hooks/hook-prompt-submit.test.ts to call
writeGitMemConfig(repoDir, { promptSubmit: { enabled: true, recordPrompts:
false, surfaceContext: true } }) (and the similar call for the other case around
lines 93-101) and remove the manual JSON write; also fix writeGitMemConfig to
deep-merge nested hook objects (instead of the current shallow merge) so that
passing an overrides object like { promptSubmit: {...} } merges into existing
hooks.sessionStart/sessionStop defaults correctly.

In `@tests/integration/hooks/helpers.ts`:
- Around line 25-38: The runHook helper currently ignores a provided cwd in its
JSON input, so update runHook to read a cwd string (e.g., const { cwd } = input
or input.cwd) and pass it into the spawnSync options (alongside input, encoding,
timeout) so the spawned process runs in the intended working directory; locate
the runHook function that calls spawnSync with TSX_BIN and CLI_PATH and add the
cwd option only when it is a valid string to avoid changing behavior when not
provided.

In `@tests/integration/hooks/hook-prompt-submit.test.ts`:
- Around line 51-63: The test "should return context in stdout" uses runHook and
currently asserts that result.stdout includes any of "validate", "input", or
"security", which is flaky; update the assertion to check for a deterministic
indicator instead (for example a known memory snippet or a stable marker
inserted by the test fixture such as a specific sentence fragment or tag), by
replacing the loose includes(...) checks on result.stdout with a single assert
that result.stdout.includes('<expected memory fragment or marker>') so the test
reliably verifies the returned context from runHook('prompt-submit', ...).