fix(io): normalize OpenDAL listed paths

[lance-gatefixer]

Summary

• route every OpenDAL-backed provider through a Lance adapter that preserves request-compatible listed paths
• decode recursive, offset, and delimiter listing results only when they mismatch the requested prefix
• preserve upstream spellings that already match, including literal percent-escape segments
• retain native start-after pushdown where supported and compare normalized paths in the bridge fallback
• add in-memory regression cases for both raw reserved characters and literal percent escapes

Root cause

object_store_opendal rebuilds OpenDAL listing entries with Path::from, which percent-encodes reserved characters. Lance constructs dataset bases with Path::from_url_path, so a raw reserved-character base can mismatch the returned listing. Unconditionally decoding the output is also unsafe because a deliberately double-encoded URI segment relies on the existing encoded listing spelling.

The adapter now preserves any upstream location that already prefix-matches the request and decodes only mismatches. This fixes raw reserved-character bases without changing literal percent-escape paths that already worked.

Validation

• cargo test -p lance-io --no-default-features --features oss object_store::opendal_store::tests::test_list_preserves_request_path_spelling
• cargo test -p lance-io --lib -- --skip uring::tests (202 passed)
• cargo check -p lance-io --all-features --tests
• cargo fmt --all -- --check
• cargo clippy --all --tests --benches -- -D warnings

The complete lance-io library test command cannot finish in this container because io_uring initialization is denied by the environment; the non-io_uring library suite passes.

Fixes #8652

[lance-gatefixer]

Blocked: the current head is held by an unrelated intermittent macOS BQ test failure that this App cannot rerun.

Remote head 1d7f8e3 contains the current main tip b945b3e. The mac-build (stable) Check failed vector::bq::storage::tests::test_accurate_approx_mode_reduces_binary_lut_quantization_error with accurate_error=0.07851982 and normal_error=0.06913757; this PR changes only lance-io, and the identical base SHA passed the macOS Check.

I fetched and verified base ancestry, inspected the live log, ran both the failed BQ test and the OpenDAL regression successfully, and pushed the assigned branch, which was already up to date. A failed-workflow rerun was requested, but the GitHub App lacks Actions rerun permission.

The smallest unblock is for a maintainer to rerun the failed mac-build (stable) job; alternatively, rerun the full Rust workflow on this head.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[lance-gatefixer]

Blocked: the current Ready PR is held by an unrelated intermittent Python nearest-neighbor test failure that this App cannot rerun.

Remote head 6b75b1a contains the current main tip cc0f085. The Python Linux 3.13 x86_64 Check (job 97195436446) failed only python/tests/torch_tests/test_bench_utils.py::test_ground_truth with one nearest-neighbor cutoff mismatch; 1 test failed and 1,379 passed. The exact base SHA passed the same Check in job 97193690776, and this PR changes only rust/lance-io with no Python diff.

I fetched and verified base ancestry, inspected the live failure log, and ran the exact test successfully with the repository uv workflow. A failed-workflow rerun was requested, but the GitHub App lacks Actions rerun permission. No attributable repair was found, so no speculative change was made.

The smallest unblock is for a maintainer to rerun the failed Python Linux 3.13 x86_64 job; alternatively, rerun the full Python workflow on this head.

[lance-gatekeeper]

✅ Gate recommendation: approve.

The conflict-free base merge leaves the reviewed lance-io adapter unchanged. Prefix-aware normalization still preserves raw reserved-character and literal percent-escape paths under the updated workspace version.