Type-based APIs become an explicit non-goal, and say so (DAP056)

[mgravell]

Rebased onto main (2026-09-11): the parity-notes commit this used to carry landed separately as #213, so it is dropped here and only the DAP056 commit remains. The parity rows are unchanged in intent — they now build on #213's "cite the generated report" framing rather than restating it.

Also picked up from #216, which landed in between: the two new TypeBasedApis goldens were written on Windows and carried \ in their [InterceptsLocation(...)]; they are normalized to / like the rest, so the fixture passes off Windows too.

Dapper's Type-argument overloads pick the row type at execution time, which is the one thing compile-time generation cannot follow. Supporting them means a build-time registry of candidate types plus runtime dispatch keyed on Type — a lookup on the hot path, an open world for the trimmer, and dispatch ILC cannot resolve unless every announced type is rooted. That is the shape #206 was closed over; rebuilding it for a lower-value feature would be incoherent.

It costs no behaviour. These already generated nothing — a five-call probe handled 0 of 5. What changes is what we say: DAP056 now fires at the call-site and names the generic overload, replacing either silence or a confusing DAP009 about an unexpected type parameter.

Detection is call-site sensitive, deliberately

GetRowParser<T>(concreteType: null) — the default, and the common case — is perfectly supportable; only a call that actually passes a Type defers the decision. So the check looks at the arguments, not the symbol. GetRowParser<T>() stays supported, and the fixture pins both halves:

handled 2 of 7 enabled call-sites (0 unsupported API, 5 refused with diagnostics, 0 skipped silently)

GetRowParser(concreteType) goes too

It is the one row where "use the generic form" is not available advice, because the choice is data-dependent — so docs/rules/DAP056.md gives the pattern that is actually correct under AOT: a switch over GetRowParser<T>() per candidate type. More code, and strictly better, because it roots exactly the types you use where a registry roots everything registered.

Reported from the generator, not the analyzer

Most of these overloads carry no sql string, so the analyzer never sees them (the reason 7 of them were in the silent bucket). SkippedSourceState gains a reason enum — plain data, so the cached model stays equatable and ModelShapeTests stays happy — and Generate emits from it.

Knock-on effects

• the surface report grows a non-goal: Type-based bucket for overloads whose Type parameter is required, so silent drops 27 → 22 and undiagnosed-unsupported 13 → 12; the mute total goes 40 → 34;
• GetRowParser's existing fixture goes from 1 silent skip to 0;
• the optional-Type overload stays in its previous bucket, because a symbol cannot say what a call will pass — the report's stated bound;
• parity.md marks the affected rows 🚫 with the decision recorded, and its "announced types" item shrinks to cover only untyped parameters: it is no longer a design that unlocks several rows.

Suites green on net8.0 (369) and net48 (362); solution builds clean.