tar: make members with duplicate slashes reachable via ls/find/glob (#1947)

[Sanjays2402]

Summary

Fixes #1947.

A tar member whose name contains redundant consecutive slashes (e.g. path/with/extra/slash//test.txt) was silently invisible to ls, find, glob and walk. As a result fsspec.open_files("tar://**/*::archive.tar") returned zero files for such archives, even though the file could be opened by its exact name.

Reproducer (from the issue)

import fsspec, tarfile

with tarfile.open("test_w_extraslash.tar", "w") as tar:
    tar.add("test.txt", arcname="path/with/extra/slash//test.txt")

of = fsspec.open_files("tar://**/*::test_w_extraslash.tar")
# before: <List of 0 OpenFile instances>
# after:  <List of 1 OpenFile instances>

Root cause

The raw member name (including the //) was used both as the self.index key and to derive the directory structure in _get_dirs. AbstractFileSystem._parent("a/b//c.txt") returns "a/b/" — with a trailing slash — because rsplit("/", 1) splits on the second slash of the //. So:

• the inferred parent-directory key became path/with/extra/slash/ (trailing slash), which never matched a listing of path/with/extra/slash, and
• the member's own computed root carried the same trailing slash.

The entry therefore fell out of every directory-traversal comparison in AbstractArchiveFileSystem and was unreachable through ls/find/glob/walk — while still openable by its exact name (which is why direct access appeared to work).

Fix

Normalise the filesystem-facing name by collapsing runs of slashes (new _collapse_slashes helper) when building both self.index and dir_cache, while retaining the original member name so tar.extractfile() still receives the exact archive key. _open normalises its lookup path too, so a member is openable by either:

• its normalised name (path/with/extra/slash/test.txt, as now returned by find/glob), or
• its original name (path/with/extra/slash//test.txt) — backwards compatible.

This matches the approach the maintainer agreed with in the issue thread ("normalizing the look-up name, while keeping the original tar key for the actual extraction").

Before / after

call	before	after
fs.find("/")	[]	['path/with/extra/slash/test.txt']
fs.glob("**/*")	['path', 'path/with', 'path/with/extra']	[..., 'path/with/extra/slash', 'path/with/extra/slash/test.txt']
fs.ls("path/with/extra/slash")	[]	['path/with/extra/slash/test.txt']
open_files("tar://**/*::…")	0 files	1 file
fs.open(".../slash//test.txt")	works	works (unchanged)

Normal single-slash archives are unaffected.

Tests

Adds test_ls_with_duplicate_slashes, parametrized over all four tar compressions (tar, tar.gz, tar.bz2, tar.xz), asserting discoverability via find/glob/ls, directory inference via isdir, and openability by both name forms.

Proven to guard the bug (source stashed, test kept):

# without the fix
FAILED fsspec/implementations/tests/test_tar.py::test_ls_with_duplicate_slashes[tar]
FAILED ...[tar-gz]  FAILED ...[tar-bz2]  FAILED ...[tar-xz]
4 failed
# with the fix
4 passed

Full local verification (py3.13):

• test_tar.py, test_archive.py, test_zip.py + memory/local/api: 328 passed, 32 skipped, 1 xfailed, no regressions.
• ruff check and ruff format --check clean on both changed files.

Diff is +76/−6 across fsspec/implementations/tar.py and its test module.

[martindurant]

Just inline this with a one-line comment

[martindurant]

tar.getnames and the normalisation happens twice here, we should probably only do it once.

[martindurant]

I think this is also iterating the same information as in _get_dirs(), please see if it could be simplified.