The half of the statifier_blocks:subchart handler that is pure in both variants: turning the document id on src into a compiled child chart, or into one of the three shared refusal reasons.

ADR-0008 decision 2 says the resolver contract is the durable module's too, "in full and without additions", and that resolution is therefore "shared code, not a second copy". This module is that shared code. StatifierBlocks.Runtime.Subchart (in-memory, Statifier.Session) and StatifierBlocks.Runtime.DurableSubchart (durable, StatifierPersistence.Driver's dispatch fun) both call resolve/3 and differ only in what they do with its two answers - which is exactly the "what differs between the variants begins after resolution has produced a chart" line the record draws.

The resolver contract itself, the behaviour, and every word of documentation about it stay on StatifierBlocks.Runtime.Subchart: this module mints nothing and is not part of the package's public surface.

:known_invoke_types is passed only when the ctx can supply it

The in-memory ctx is a Statifier.Invoke.Handler.ctx/0 and carries :invoke_handlers, the set the runtime will classify against, so a child compile gets ADR-0004 decision 8's opt-in lint. A durable dispatch context carries no such key - the durable seam's registry is the host's dispatch fun, not a handler map - so the option is omitted rather than passed as an empty set. Omitted, the compiler does not run the lint at all (StatifierBlocks.Compiler's Keyword.fetch/2). Passed empty, it would run and report every invoke type in the child as unregistered - starting with statifier_blocks:subchart itself in a nested durable subchart (ADR-0008 decision 6) - which is a false claim about the host's registration rather than a true one. The lint is warning-only (StatifierBlocks.Compiler.InvokeTypes), so neither choice can change a refusal; the choice is between silence and noise, and silence is the honest one where there is nothing to compare against.