perf: stream prune integration checks, parallel removals

[max-sixty]

Replace the blocking par_iter().collect() barrier in step_prune with channel-based streaming. Integration checks run on a background thread via rayon, sending results through a crossbeam channel as they complete. The main thread processes results as they arrive for age filtering and candidate collection, overlapping with still-running checks.

Phase 2 removes confirmed candidates in parallel via into_par_iter(). Candidates are sorted by original check order for deterministic output.

Performance (100 BranchOnly removals, release build):

Metric	Before	After
Non-dry-run (100 branches)	7.8s	2.7s
CPU utilization	139%	485%

Tests with multiple removal candidates pin RAYON_NUM_THREADS=1 for deterministic snapshot output.

This was written by Claude Code on behalf of @max-sixty

codecov/patch failed (94.8% — 4 uncovered lines in new code)

[worktrunk-bot]

codecov/patch failed — 94.8% patch coverage (4 uncovered lines of 77).

The code itself looks good (clean channel pattern, correct deterministic ordering, proper cleanup). The uncovered lines are:

1. Progress message (L1549–1552): Behind check_items.len() > 10 — no test creates that many branches. This is UX-only, no logic.

2. drop(tx) (L1578): Empty check_items path. Would need a test with a repo that has only the default branch and no other worktrees/branches.

3. Non-dry-run skipped_young report (L1744–1750): The refactoring duplicated the "Skipped X (younger than Y)" message into both dry-run and non-dry-run branches. The existing test_prune_min_age_skips_young and test_prune_stale_plus_young tests only exercise dry-run. A non-dry-run variant with --yes --min-age=1h and a young worktree would cover this.

The easiest fix for coverage is #3 — add a --yes (non-dry-run) variant of one of the existing min-age tests.

CI failed — snapshot mismatch in test_prune_removes_integrated_detached

[worktrunk-bot]

Linux CI failed: test_prune_removes_integrated_detached snapshot mismatch (log).

This test produces two parallel candidates (the detached worktree + the orphan detached-integrated branch), but doesn't pin RAYON_NUM_THREADS=1. The branch deletion finishes before the worktree removal, flipping the output order. Other multi-candidate tests (lines 144, 224, 288, 324) already pin it.

Fix in tests/integration_tests/step_prune.rs at line 197 — same pattern as test_prune_multiple:

    let mut cmd = make_snapshot_cmd(
        &repo,
        "step",
        &["prune", "--yes", "--min-age=0s"],
        None
    );
    cmd.env("RAYON_NUM_THREADS", "1"); // deterministic output order
    assert_cmd_snapshot!(cmd);

Then cargo insta test --accept to update the snapshot. I can push this fix if you'd like.