pod bootstrap sparse-cone excludes tracked data/ inputs → fresh-pod FileNotFoundError (from #2203 Phase 3)

Goal

Close the pod-bootstrap sparse-cone gap that makes an experiment crash on a fresh pod when it reads a git-TRACKED input under data/ — decide and implement the right fix (extend the default bootstrap cone to cover small tracked data/ input subdirs, OR a documented mechanism for a plan to declare such a cone, OR make experiments stage these inputs from HF), and document it so planners/experimenters expect it.

Context (incident, #2203 Phase 3, 2026-08-10)

bootstrap_pod.sh clones with --filter=blob:none + cone sparse-checkout; the realized cone on a fresh pod-2203 (H200) was configs docs eval_results/issue_2203 figures/issue_2203 scripts src tests — data/ is NOT in the cone. Phase 3 (scripts/issue2203_phase3.py) crashed at build_jailbreak_set → load_role_list() with FileNotFoundError: data/assistant_axis/role_list.json, AFTER the 32B model loaded (~1 GPU-cycle + a crash-fix round wasted). The file IS tracked on the branch and NOT gitignored (git ls-files returns it, git check-ignore empty) — but its blob is never materialized because data/assistant_axis/ sits outside the sparse cone. Recovery was git sparse-checkout add data/assistant_axis on the pod (materialized role_list.json + extraction_questions.jsonl + 276 instruction files, byte-integrity verified).

Why the existing guard did not catch it: the #1739 gotcha (.claude/rules/gotchas.md, "Partial-clone pods") covers a pod reading ANOTHER issue's committed artifacts via BOOTSTRAP_EXTRA_CONES — it does NOT obviously cover THIS repo's own tracked data/ inputs excluded by the DEFAULT cone. data/ legitimately holds both huge re-downloadable caches (data/issue_<N>/hf_dl) and small tracked inputs (data/assistant_axis/), so the fix must exclude the caches while including the tracked inputs.

Note: Phase 0/1/2 ran on a DIFFERENTLY-bootstrapped pod (the first pod-2203, full bootstrap) and did not hit this — the interrupted-bootstrap + manual-git checkout issue-2203 path on the Phase-3 pod is what surfaced the cone gap, but the gap is latent for any fresh pod whose experiment reads a tracked data/ input.

Acceptance

1. A decided fix implemented (one of: bootstrap_pod.sh default cone includes tracked data/** input subdirs excluding the hf_dl/g*_dl cache dirs; OR a BOOTSTRAP_EXTRA_CONES-style declaration path documented + wired for data/ inputs; OR an experiment-side stage-from-HF convention for data/assistant_axis-class inputs).
2. The relevant gotcha / planner / experimenter surface documents the expectation so a planner declares (or an experimenter stages) tracked data/ inputs for pod phases.
3. Any mechanical check (e.g. tests/test_bootstrap_pod_issue_cones.py) updated if the default cone changes.