L2 gets an exit criterion, because the measurement gave it one

[milyin]

Docs only. Rewrites the L2 section of docs/language-integration.md, which #229's body mirrors.

The section was a by-file checklist written before the stage was investigated, and it inherited the map's default assumption — that some sites stay because they inspect types the adapter synthesized.

The measurement

Instrumented Registry::ensure_entry (the one place a type-table cell is born) across example-cbindgen, covertest-kotlin, perftest-c and perftest-kotlin. TypeSubject::Adapter has two members:

Cell	Where it came from
MaybeUninit<Payload>	source writes &mut MaybeUninit<Payload>; the scan peels the &mut syntactically. The model absorbed MaybeUninit into RefMode::Out, so it has no node for the peeled spelling — and the cell resolves to no converter
Option<Summary>	source writes Option<&Summary>; the peel family strips the & inside the Option and re-composes a spelling nothing wrote

Neither is a type a binding invented. Both are manufactured by the syntactic peeling L2 deletes, and both underlying types are ordinary source types.

Corroborating that the category has no genuine members: unfold.rs, expand.rs and registry/walk.rs contain zero mentions of TypeEntry or a wire type, so the "legitimately the adapter's business" exemption has no members in this stage — all 45 sites classify source types. The .cross() path, which the code comments describe as the adapter-authored one, is exercised by covertest-kotlin and produced no cells either.

Measured in-tree only, stated as such in the doc — the same shape of claim #248 makes about fallible_parts logging zero. zenoh-flat's consumers are out of tree and do not currently build.

What the section now says

An exit criterion ahead of the checklist, because it is stronger than a count and is what each PR reports against:

TypeSubject::Adapter reaches zero members, and the variant is deleted.

TypeCell's subject collapses to the TypeRef; location() becomes origin.location gated by the existing has_position(), and kind() — already #[cfg(test)] — goes with it. Two production readers, both in resolve.rs.

Then five PRs summing to exactly 45:

	Owns	Sites
L2a	the scan walks model edges	13
L2b	one layer read	26
L2c	the enum shape and its numbers	2
L2d	canonicalization moves home	4
L2e	delete TypeSubject::Adapter	0

Each says whether it migrates a site or deletes the code holding it — L2a drops three dead arms (lower_type refuses tuples and raw pointers; Group/Paren are transparent; named() drops generic args by design), and L2d is labelled a move, not a migration, the same distinction #248's deletions get.

L2b and L2c also close two of the ledger header's listed blind spots by removing the code: the seg.ident == "Option" family, and enum_shape / first_payload_variant, which take a syn::ItemEnum that WATCHED cannot see.

Not changed

No counts move — this records the plan, not progress. The #248 is deletion, not migration note stays, and the older TypeKey-from-a-TypeRef bullet is folded into L2b rather than dropped.

[milyin]

Closing unmerged: the stage plan belongs in the umbrella PR description and in #211, not in a file in the repo.

The measurement and the L2a–L2e split it produced are carried over to #229's body and to #211's text, so nothing in it is lost. The two corrections it also carried — the ledger standing at 167 rather than 202, and the caller counts that make L2 end at 10 sites rather than 0 — go with them.

Superseded by the description updates on #229 and #211.