The truth readers open a zarr without the guard written for it #231

tabascal/truth.py has three readers that take a path and open it bare:

def read_true_vis_ast(zarr_path, data_col=None):   # truth.py:144
    xds = xr.open_zarr(zarr_path)
def read_true_vis_rfi(zarr_path, data_col=None):   # truth.py:153
    xds = xr.open_zarr(zarr_path)
def read_true_gains(zarr_path):                    # truth.py:162
    xds = xr.open_zarr(zarr_path)

A store that is missing, unreadable or corrupt gives whatever xarray raises, from inside component setup, without naming the configuration option that asked for the truth.

There was a helper for exactly this

#230 removed _open_zarr, which was dead:

def _open_zarr(config: Dict):
    """Open the sim zarr, raising a clear TruthError if it cannot be read."""
    path = _zarr_path(config)
    if not path or not os.path.exists(path):
        raise TruthError(f"No tab-sim truth available: simulation zarr not found at {path!r}.")
    try:
        return xr.open_zarr(path)
    except Exception as e:
        raise TruthError(f"Could not open simulation zarr at {path!r}: {e}") from e

It took a config; the readers take a path, so it could never have been their caller. It was written for this and never wired in, which is why it read as dead to a reachability sweep. Deleting it was right; the gap it was meant to close is still there.

Why it is not simply reachable today

require_truth runs before TabConfig setup and fails fast when a config needs truth the zarr does not provide — it checks the path exists and that the required variables are present. So the common cases are caught early and well.

What is not caught: a store that exists and passes available_truth but fails on read — truncated, permissions, a partially written zarr, a store on a filesystem that goes away between the preflight and the read. available_truth opens it in a try that returns all-False on failure, so an unreadable store makes require_truth report the truth as missing, which is a reasonable message; but a store that opens for the preflight and fails later reaches the readers bare.

What to do

• Give the three readers the same treatment: a TruthError naming the path and the underlying error. A shared path-taking helper rather than the config-taking one, so the readers can actually use it.
• Decide whether available_truth's silent except Exception: return flags should distinguish "no store" from "store I could not read" — right now a corrupt zarr is reported to the user as an absent one, and those want different fixes.
• Worth a test with a deliberately truncated store, which nothing currently covers.

Related

• Delete unreachable code, and the config keys advertising it #230 — where _open_zarr was removed and this gap noticed