A specialization minted in a unit's implementation blocked the importer's own

Filed as a record rather than as work: found, diagnosed and fixed inside one session, and it is on origin for a few hours between two commits. It gets a ticket instead of a bare logbook line because the code comments explaining three separate edits name it, and because the failure mode is one anybody adding a ninth declaration table can repeat.

Repro

ugimpa declares TBox<T> in its interface. ugimpb specializes TBox<Integer> from an IMPLEMENTATION uses and publishes only an Integer. A program then uses both and writes var b: TBox<Integer>.

pascal26:3: error: unknown type: TBox$Integer

FPC 3.2.2 and the pinned binary both print 42 202 4.

The seam

Two commits landed the same afternoon and neither is wrong:

• Track D's 84ea8e470 stamped every declaration table with the section its rows were declared in, so FindUClass correctly refuses ugimpb's private TBox$Integer class row to an importer.
• This slice's 01998adb8 taught ParseSpecialization that "an exact re-statement is a no-op" is a question about the TEMPLATE INDEX, not a template name.

What broke is the check between them. FindSpecialization asks DeclVisible(SpecUnitIdx[i]) — the UNIT half of the answer and not the section half — so:

The permissive check suppressed the work the strict check then demanded, so the program's own declaration was never emitted. That is worse than either rule applied alone, and it is the general shape: a visibility question answered in two places will eventually be answered two ways.

Fix

Specializations[] gained SpecDeclImpl[], stamped from DeclInImplNow at registration exactly as the other eight tables are, plus an IMPLTAB_SPEC tag so p.implleak names the right row. FindSpecialization calls DeclVisibleSect. ImplPrivateApplies treats the new tag like IMPLTAB_UCLS — a minted specialization name has no builtin chain to outrank, so the ambient exemption applies to it, which is the right answer for the same reason it is right for a class.

Measured

Test: test/test_generic_implsect_dup.pas + generic_implsect_units/.

The other thing this run turned up, filed separately

frankD asked for the one case they had no coverage for: a generic TEMPLATE declared in a unit's implementation section, named by an importer. pxx accepts it and FPC refuses it, and the pinned binary accepts it too — pre-existing, not fallout from the boundary work, because Templates[] is a token-arena registry with no unit or section channel for a visibility answer to live in. [[bug-p-a-generic-template-declared-in-a-units-implementation-is-visible-to-its-importers]], prio 35.