compass-calibrate-formatlisted

# compass-calibrate-format When words keep missing the target shape, switch to examples. The guard enforces the example-count band and the diversity rule. `GUARD="python3 ${CLAUDE_PLUGIN_ROOT}/scripts/compass.py"` ## Assemble the example set - **Source examples from real prior outputs first.** Only construct examples if no real outputs exist. - Use **2-5 examples total.** Fewer than 2 is not few-shot anchoring; more than 5 usually means the format is still underspecified — tighten it instead. - **Do not use all-alike examples.** At least one must sit **near a boundary** between plausible interpretations, so the examples actually pin the decision. - If you must **construct** the set, include a **happy path** and **at least one edge case**, and mark the boundary example with `near_boundary: true` and each example's `kind` (`happy-path` / `edge-case`). ## Validate Real examples: ``` echo '{"constructed":false,"examples":[{"in":"x","out":"y"},{"in":"a","out":"b"}]}' | $GUARD calibrate - ``` Constructed examples (diversity enforced): ``` echo '{"constructed":true,"examples":[ {"kind":"happy-path","in":"normal","out":"…","near_boundary":false}, {"kind":"edge-case","in":"empty","out":"…","near_boundary":true} ]}' | $GUARD calibrate - ``` A non-zero exit means too few/many examples, or a constructed set missing a happy path, an edge case, or a boundary example — fix before producing the artifact. ## Output - a formatted artifact matching the decision boundary the