experimenter input-data gate: probe sparse-checkout before planned-input-data-missing (#2388 Pod B lesson)

Goal

Close the experimenter-spec gap surfaced during the #2388 Pod B launch (2026-08-20): the § "Before Running" input-data completeness gate reads a missing git-resident file as planned-input-data-missing and would post epm:failure + refuse launch, but on a SPARSE pod checkout (bootstrap clones scoped to the current issue's cones, e.g. eval_results/issue_2388 only) banked inputs from a PARENT/SIBLING issue (eval_results/issue_1739/...) are absent on disk while TRACKED at HEAD. A respawn hits the identical sparse clone, so the failure loops.

Proposed change

In .claude/agents/experimenter.md § "Before Running" (input-data completeness gate): before posting planned-input-data-missing-on-pod for a git-resident path, probe git sparse-checkout list + git ls-tree HEAD <path>; if the path is tracked but outside the sparse cones, run git sparse-checkout add <needed dirs> and re-stat instead of failing. Mirror the same probe in any bootstrap/staging doc that promises "the repo clone has the tracked tree" (.claude/rules/gotchas.md candidate entry).

Provenance

workflow_fix_target: .claude/agents/experimenter.md Surfaced by the #2388 Pod B experimenter (session cmt0xeq8sqpy3xw0uefy8ar19 dispatch, 2026-08-20T14:26Z), which self-recovered by adding 4 sparse cones and verified all 6 banked inputs on disk (4.6-30.9 MB). failure-lesson block carried in the agent's final report; root_cause_confirmed: yes.