Upstream refresh coverage

Requested outcome	Evidence condition	Current scope / remaining gap
Avoid duplicating upstream cache/telemetry work	Pinned full upstream issue/PR review plus source comparison	Closed reviews selected; as-observed cutoff retained, no claim of perpetual freshness
Reuse retained signed GLM package	Exact template compatibility provenance and independently checked rendering/boundary behavior	Closed compatibility evidence selected; no new preparation or graph family
Current-upstream compatible baseline	Frozen image/recipe/package identities, all requested phases and independent accounting	Closed: 391 adapter requests plus one smoke; all six incarnations retained
Fixed1 repeatability/cache behavior	Three complete fresh-incarnation repetitions with raw per-request output/work and metrics	Closed: three 52-request reverse replays, each with ten host fetches; no old pressure-counter synthesis
Protected small12/pageheavy behavior	Complete frozen warmup/prime/measured populations and gate evaluation	Closed: 24 small12 and 12 pageheavy measured requests; separate warm GPU-hit populations, no code-only cross-driver comparison
Failures and exact recovery	Every predecessor plus final canonical restoration, direct/routed checks, readmission and monitoring	All predecessors retained; canonical exact restoration passed with no recovery errors
Numerical equivalence and promotion	Separately declared numerical/quality gates and human authorization	Not established by this campaign; no promotion or default change
Complete publication	Closed inventory, sanitized native artifacts, deterministic manifest, privacy and claim reviews, full documentation matrix	Final bundle is separately validated; hosting is a later owner-controlled action

Completed canonical population: 192 measured, 180 prime, 15 warmup and four fixtures, plus one smoke. Failed predecessors are retained separately and not erased from campaign history. Missing captures remain missing; neither empty fields nor shared absence proves matching work.