Generalize the cross-instance type check into a portability matrix

[candrewlee14]

Stacked on #6200. Until that lands this branch contains its commit too — review only Generalize the cross-instance type check into a portability matrix. I will rebase once the parent merges.

What

#6200 fixes SQL and guards it with a single hardcoded assignment. That guards exactly one type. This replaces it with a table.

Why

The underlying bug is not specific to SQL: any class whose private/protected members survive into the emitted .d.ts is compared nominally, and so is non-portable between two installs of the package. That is a property of the whole published surface, and it moves in both directions — a refactor can introduce a leak, and an unrelated @internal can quietly remove one.

A single assertion catches neither. The matrix asserts both:

• a type recorded as portable that stops being portable fails (regression);
• a type recorded as a known leak that becomes portable also fails, telling you to flip the flag.

Shape

PORTABILITY_MATRIX lists each probed type with its subpath, any needed type arguments, and whether it currently survives crossing between installs. Every portable: false entry carries a blockedBy naming the member responsible, so the table doubles as the work queue. All probes compile in one tsc program — the whole test runs in ~1.4s, comparable to the single assertion it replaces.

The recorded state is measured, not aspirational: it is exactly what the current published declarations do.

Follow-ups

Later PRs in this stack flip rows to portable: true as the underlying members are stripped.

[candrewlee14]

Consolidated into #6200. The whole change is 244 lines across 16 files, and because a fork PR cannot base itself on another fork branch, every PR in the stack was already showing all of its ancestors' commits — so the split bought no review isolation, just six merges and five rebases. It is now one PR with these commits intact, plus #6205 for the single change that touches observable API.