Select resolver prefetch candidates in one pass

[charliermarsh]

Summary

Batch prefetch currently selects up to 50 candidates by repeatedly invoking the full candidate selector and intersecting a singleton complement into the remaining range. On backtracking-heavy packages, those repeated traversals and increasingly fragmented ranges add meaningful resolver CPU before the metadata requests are even issued.

This adds a batched selector that visits each version map once, preserves highest/lowest ordering, prerelease and cutoff handling, index strategy and duplicate-index precedence, compatible/incompatible candidates, and the compatible-to-in-order phase transition, then enqueues the same eligible wheel metadata requests. This is distinct from #20465: it does not suppress later prefetch opportunities; it only makes selection within each batch linear.

This is deliberately a draft because collecting a batch before enqueueing its first request can trade away some cold-network overlap. The target benchmark is warm/offline.

Performance

Measured directly against latest main (bedddb53a) with profiling binaries and a prewarmed PyPI cache. All runs were offline uv pip compile, Python 3.12, fixed UV_EXCLUDE_NEWER=2024-08-08T00:00:00Z, UV_CONCURRENT_CACHE_READS=4, CPUs 8-15, 20 warmups, and 80 alternating pairs. Every generated requirements file matched byte-for-byte.

workload	wall before → after	wall change (95% CI)	CPU before → after	CPU change (95% CI)
jupyter (97 packages)	97.6 → 98.5 ms	+4.8% [-0.3, +10.2]	54.6 → 54.5 ms	+0.3% [-0.6, +1.2]
boto3 (7 packages, backtracking)	163.3 → 157.3 ms	-3.6% [-6.6, -0.5]	229.2 → 218.8 ms	-6.6% [-9.1, -4.1]
airflow (559 packages)	243.4 → 241.8 ms	-0.2% [-2.0, +1.5]	370.6 → 369.3 ms	-0.7% [-1.7, +0.2]
large-index (26 large Simple pages, no deps)	102.6 → 103.0 ms	+1.4% [-3.6, +6.3]	58.5 → 58.3 ms	-0.3% [-1.1, +0.5]

boto3 is the meaningful win, which matches its roughly 94 large prefetch batches. A focused selector-equivalence matrix covers highest/lowest, all three index strategies, prerelease modes, duplicate versions, compatible/incompatible/cutoff entries, and segmented ranges; it also matched 300,000 randomized cases against repeated selection. The full resolver unit suite (65 tests), strict clippy including tests, formatting, and diff checks pass.

[astral-sh-bot]

uv test inventory changes

This PR changes the tests when compared with the main base revision.

• Added tests: 1
• Removed tests: 0
• Changed suites: 1

uv-resolver: +1 / -0

Added:

• uv-resolver::candidate_selector::tests::batched_selector_matches_repeated_selection

Removed: none