Sparse-checkout issue worktrees: cut per-worktree disk cost ~14G → ~2G

Goal

Every /issue worktree (Step 4a) and agent-isolation worktree should materialize only the directories issue work actually edits, so 40+ concurrent worktrees cost ~80G instead of ~560G.

Background (incident 2026-06-11)

VM root disk hit 100% (1.2 GiB free) with 59 worktrees present. Measured: each worktree is a full ~14G checkout, of which eval_results/ = 11G and external/ = 2G — history an issue branch essentially never needs materialized. The stale-worktree audit reaps idle ones, but with ~27 parallel sessions the steady-state worktree count stays high; reducing per-worktree cost is the structural lever (the watcher-side auto-remediation fix for the cadence gap is being applied separately).

Proposal

Create worktrees with git worktree add --no-checkout, then enable cone-mode sparse-checkout excluding eval_results/, external/, and (optionally) ood_eval_results/ + .arxiv-papers/ before the first checkout. New files under eval_results/issue_<N>/ can still be created and committed from a sparse worktree (the sparse cone gates materialization of existing tracked files, not the addition of new ones) — verify this explicitly for the analyzer's figure/eval-JSON commit path.

Open questions the plan must settle

1. Which paths the analyzer / experiment-implementer / upload-verifier actually READ from the worktree (vs repo root): parent-issue eval JSONs for comparison plots are the risk case. Decide: read those via repo root (task.py find style helpers), or sparse-add the specific parent dir on demand (git sparse-checkout add eval_results/issue_<M>).
2. Whether the Step 10d merge guards + surgical-checkout path interact with sparse cones (the additive-checkout file list includes eval_results/issue_<N>/ paths).
3. Whether worktree_audit.py's uncommitted-changes guard behaves identically under sparse checkout.
4. Migration: existing worktrees stay as-is (reaped naturally); only new creations are sparse.

Acceptance criteria

1. /issue Step 4a (SKILL.md) + any helper that cuts worktrees documents and uses the sparse recipe.
2. A fresh issue worktree measures ≤2.5G.
3. End-to-end: an issue cycle (implement → commit figures + eval_results/issue_N → merge) completes from a sparse worktree with no path errors.
4. Worktree audit + Step 10d guards unaffected (tests).

Notes

• kind: infra. Touches .claude/skills/issue/SKILL.md (public worktree contract) — flagged greenlight-class per the workflow-fix protocol, hence a proposed task rather than an auto-applied fix.