Tests fail due to locale-dependent smart quotes in error message matching (`fixed = TRUE`)

[villegar]

Summary

Some tests fail when running with Armadillo (and certain R/locale configurations) because error messages use Unicode smart quotes (‘ ’) instead of ASCII single quotes ('), causing strict expect_error(..., fixed = TRUE) comparisons to fail even though the correct error is thrown.

Failing Test Example

expect_error(
  datashield.aggregate(conns = NULL, expr = calltext),
  "unable to find an inherited method for function 'dsIsAsync' for signature 'conn = \"NULL\"'",
  fixed = TRUE
)

Observed failure:

• Expected: 'dsIsAsync'
• Actual: ‘dsIsAsync’

Error message received:

unable to find an inherited method for function ‘dsIsAsync’ for signature ‘conn = "NULL"’

Root Cause

The difference is purely typographical:

• ASCII quotes: '
• Unicode smart quotes: ‘ ’

This variation depends on:

• R version
• system locale (e.g. UTF-8 vs C)
• platform/container environment
• upstream package message formatting (S4 dispatch errors)

Since the test uses fixed = TRUE, it performs an exact byte match and fails despite the semantic message being identical.

Why This Is Problematic

1. Makes tests locale-dependent and non-portable
2. Causes false negatives in CI and Docker environments
3. Fails on systems that default to UTF-8 smart quotes
4. The underlying functionality is correct, but the test is overly strict

Suggested Fix

Replace strict fixed-string matching with a more robust pattern, for example:

expect_error(
  datashield.aggregate(conns = NULL, expr = calltext),
  "dsIsAsync"
)

or remove fixed = TRUE and match a stable substring.

Alternatively:

• Normalise quotes before comparison, or
• Avoid matching the full error string from S4 dispatch (which is not guaranteed to be stable across R builds)

Environment Where Observed

• Backend: ArmadilloDriver
• Docker-based test environment
• UTF-8 locale
• R producing Unicode curly quotes in S4 error messages

Workaround

Setting locale to C avoids the issue:

Sys.setlocale("LC_ALL", "C")

However, tests should ideally be locale-agnostic rather than requiring environment-specific configuration.

[StuartWheater]

This behavior has been observed, but root cause was not known.

[StuartWheater]

My concern is if error messages don't have well defined predictable values programmatic fault-tolerance won't be possible, a such code won't be able to determine, accurately, if error has occurred.