Harden eval-sensitive tests: retry racy group kill, hermetic worktree fixture

[w4ffl35]

Two more RSI baseline blockers found by reading the fleet eval artifacts (run ...00abea81 got to 118/194 files):\n\n1. muse-executor group-kill EPERM: kill(-pgid) can throw for a transient group under load (one strike in a loaded eval sandbox, 10/10 clean container re-runs). Retry briefly with backoff; the muse-is-dead invariant below still fails loudly if the worker survives a signalled group.\n2. run-tests-script worktree test requires the ambient checkout to be a git repo (git worktree add), but eval snapshots extract files without .git — deterministic failure when reached. The fixture now builds its own throwaway parent repo (git init) with an untracked node_modules symlink, preserving the exact walk-up scenario and no longer mutating the real repo's worktree list.\n\nBoth verified locally (5/5 and 12/12) plus tsc clean; full-suite container repro in progress.

[w4ffl35]

Note: PR title/body edits are failing (GitHub GraphQL projects error), so the title is stale — the branch now also includes fix 3 (openshell-provider reap TOCTOU: treat a stat-vanished-between-probe-and-read as terminal, same as ESRCH). Commit message is accurate. All gates green locally.