fix(dataset_tools): don't split XRootD URLs at the port colon when parsing a files list

[NJManganelli]

Part of #1578 — the filespec.py rsplit(":") parsing bug.

When DatasetSpec is given files as a list, each entry was split with rsplit(":", maxsplit=1) to peel off a trailing ROOT object path, guarding only remainders starting with //. This mishandles XRootD URLs that carry a port, e.g. root://host:1094//store/f.root: the split lands on the port colon, silently mangling the filename to root://host and inventing an object path 1094//store/f.root.

This PR delegates the split to uproot's battle-tested file_object_path_split (via a small _file_object_path_split helper with a semantics-mirroring pure-Python fallback should that internal helper move), which correctly understands the root:// scheme, ports, //-paths, object paths containing /, and local absolute/relative paths. The dict-input path and the dual dict/model contract are unchanged. Regression tests cover the full case matrix (file.root:Events, file.root:Dir/Tree, scheme URLs with and without ports and object paths, local paths); the port-URL case fails on master and passes here. tests/test_dataset_tools_filespec.py: 152 passed (141 existing + 11 new). Pre-commit clean.

🤖 Generated with Claude Code

[lgray]

🤖 AI text below 🤖

Summary. This correctly fixes the port-colon mis-split for files-list input by delegating to uproot's file_object_path_split. I verified the fix end-to-end: on master a list entry like root://host.cern.ch:1094//store/data/dir.parquet is mangled into filename root://host.cern.ch + object_path 1094//... (→ RuntimeError in identify_file_format); at PR head it parses correctly. The uproot-backed helper matches uproot exactly across the full matrix (port ± object path, object paths with /, local abs/rel, https://host:port, Windows drive). The dict-input path is genuinely untouched (split runs only in the isinstance(files, list) branch of the mode="before" validator), and dict/model round-tripping still passes. Full tests/test_dataset_tools_filespec.py = 152 passed. Good, well-scoped fix — one real issue in the fallback, below.

Should fix

• The pure-Python fallback re-introduces the exact bug this PR fixes, and nothing tests it (src/coffea/dataset_tools/filespec.py:38-44). The fallback splits on rpartition(":"); for a port URL without an object path the last colon is the port colon, so root://host:1094//store/data/dir.parquet → ('root://host', '1094//store/data/dir.parquet') — precisely the mangling master produced. I confirmed this with a differential test (uproot vs fallback over the PR's own matrix): they disagree on all three port-URL-no-object-path cases, and a forced-fallback run (del uproot._util.file_object_path_split) fails 2 of the 8 helper cases — including the parquet-directory case that test_list_input_xrootd_url_with_port_no_object_path is meant to guard. Because uproot's splitter keys off a .root/regex match while the fallback treats any non-/ trailing colon-suffix as an object path, the fallback can't be made faithful with a one-line tweak. Options, in order of laziness:
  • Drop the fallback entirely and import from uproot._util import file_object_path_split at module top. coffea already hard-depends on uproot; if that private symbol ever moves, a loud ImportError at import time is far better than silently reverting to buggy parsing. This is the smallest, safest diff.
  • If you keep a fallback, it must reproduce uproot's .root-anchored semantics (not rpartition), and add a test that runs the regression matrix with the fallback forced active — otherwise the safety net is untested and wrong for the very case it's insuring.
• except Exception is too broad (filespec.py:37). The comment says the guard is for uproot's helper being "unavailable/moved," i.e. ImportError/AttributeError. As written it also swallows any exception uproot's splitter itself raises on some input and silently falls through to the (buggy) fallback. Narrow to except (ImportError, AttributeError).

Nits / Optional

• files_list = files (filespec.py:514) is a redundant alias; you can iterate files before rebinding, or just inline. Trivial.
• Worth a one-line comment noting uproot._util.file_object_path_split is a private uproot API (hence the guard) — future readers will wonder why a public function isn't used; there isn't one.

Test coverage. Non-vacuous and discriminating — proven by revert: test_list_input_xrootd_url_with_port_no_object_path fails on master with the mangled root://host.cern.ch parse and passes here, and the tests assert the full (filename, object_path) pair, not just "didn't raise." The matrix is well chosen (port ± object path, object paths with /, local abs/rel, bare filename). One gap: every test exercises only the uproot path; the fallback (the whole reason for the try/except) is never run, and it fails the port case when it is — please add a forced-fallback test or delete the fallback per above. A C:\...:Events case and an explicit dict-input-unchanged assertion would round it out (both verified fine here).

Pydantic. Idiomatic: the split lives in the existing @model_validator(mode="before"), correct mode for reshaping raw input, mutating-and-returning data on the list branch only; no mutable-default or recompilation concerns introduced.

[NJManganelli]

Rather than have a fallback, just rely on uproot's implementation, and add a test to catch any functionality drift. Rerunning tests

[NJManganelli]

This PR straightforwardly delegates to a better implementation in uproot, and we add a few tests to try to ensure that functionality suits our needs. Anything else needed?