Credential guard tracebacks on a non-integer POSTGRES_PORT instead of reporting it #92

Problem

scripts/check_deploy_credentials.py promises "one line per problem" and reports malformed DATABASE_URL / DATABASE_READONLY_URL / REDIS_URL values instead of raising (see PR #90 / #90 review fix 713735b). The same guard still crashes with a traceback when the datastore is configured via components and POSTGRES_PORT is not an integer:

export POSTGRES_PASSWORD=app_password_9f3a
export POSTGRES_READONLY_PASSWORD=ro_password_7c21
export REDIS_PASSWORD=cache_password_4b18
export POSTGRES_PORT=abc            # typo
unset DATABASE_URL DATABASE_READONLY_URL

python scripts/check_deploy_credentials.py
# ValueError: POSTGRES_PORT must be an integer, got 'abc'.

audit_credentials() calls agentkit.config.resolve_postgres_url() (twice) without guarding it, and that function raises ValueError for a non-integer POSTGRES_PORT (agentkit/config.py, documented under "Raises"). Both the application (Settings) and the guard hit the same exception; the result is a traceback in the compose migrations service instead of the actionable problem line the rest of the audit produces.

Impact

• Severity: low. The deployment still fails closed — the migration service stops — so no insecure configuration can slip through; only the error presentation is wrong.
• Introduced by PR fix(audit): remediate all 18 AUDIT findings as a ten-branch stack #84 (75f14d1), i.e. pre-existing; out of scope for PR fix(audit): close the five PR #84 review follow-ups as a five-layer stack #90 and noted in the post-merge review.

Suggested fix

Wrap the two resolve_postgres_url() calls in audit_credentials() with a small helper that catches the ValueError and appends POSTGRES_PORT is not a valid integer. to problems, in the same style as _parse_url() / _resolve_redis_url(). Settings keeps its fail-fast behavior (a ValidationError at startup is fine there); only the guard needs the stable message. Add a regression test with POSTGRES_PORT=abc in component mode.

Relates to #86 (guard correctness follow-ups from the PR #84 review).