CW-ADR-0007: Phantom combos -- a curated denylist of upstream gaps

Status: accepted (direction: make it data, then self-maintaining)

Decision

A curated denylist of upstream gaps, subtracted at matrix time. Not build-and-fail discovery on every run (red noise, wasted CI, confusing late failures at torch-install time); known-impossible cells cost zero.

PHANTOM_COMBOS: an explicit denylist of (cuda, torch, python, platform) tuples known to be unpublished upstream. Matrix generation silently drops matching cells. Curation is manual: when a new combo lands, missing upstream cells are discovered by build failure and added.

Context

Upstream PyTorch does not publish every (cuda, torch, python, platform) cell -- e.g. several cu129 Windows cells for torch 2.10/2.11 simply never shipped. A grid cell whose torch wheel does not exist fails late and confusingly: the matrix generates, the job spins up, and dies at uv pip install torch==... inside the build-env setup. Nothing at matrix-generation time validates a cell against upstream reality.

Consequences

• Known-impossible cells cost zero CI and produce zero red noise.
• Verified defect (2026-08 audit): the list is hardcoded in TWO scripts (generate_matrix.py and gap_analysis.py) and the copies have already drifted -- the gap-analysis copy is missing entries, so its reports overcount gaps.
• Direction (agreed): move the list to one data file (scripts/phantom_combos.json) read by both scripts; then let the upstream watcher (CW-ADR-0008) maintain it automatically -- the same fetch that detects new combos also knows exactly which (python, platform) cells upstream did not publish, turning phantom curation from fail-then-edit into data derived at detection time.