The merge of five branches surfaced a regression none of them could see alone:
eleven `money_*` / `tag_newtype_*` cells began reporting E0742 on correct code.

`Money<USD>` — where `USD` is a nullary variant of the parameter's declared
`data Currency` — is a concrete TAG, not an unbound name. It broke because the
binder slice closed the head-resolution blind spot (a stdlib head missed in an
app module, so the whole value-slot check was silently skipped and a field
absorbed to unbounded) by giving `resolve_head_type_id` a project-roster
fallback after module scope. `resolve_named_type` did not get the same
fallback. So the HEAD `Money` resolved globally while the PARAMETER's declared
`Currency` still stopped at scope, `value_slot_tag` could not see `USD` as a
variant, and a legal tag was diagnosed as a free binder.

Closing the head hole was right; asking for two names two different ways was
not. `resolve_named_type` now ends in the same roster lookup, so what CHECKS
and what LIFTS cannot tell different stories about one program — which is the
invariant `param_args`' module doc already claimed and no longer had.

Measured after the fix, not argued:

  - `Money<USD>`                     clean (was E0742)
  - `List<byte, n>`, n free          E0742  — the free-binder rule holds
  - `List<byte, WOBBLY>`, unfoldable E0742  — the head blind spot stays closed
  - `Pair2<1 + 1, Int>`              clean  — the lead defect stays fixed

Also: Grammar.ckx's M1/M2 sites moved (List.ck:272→285, 222→235). Values are
unchanged at 3 and 4; the binder slice added `<n: Size>` to five signatures and
shifted the lines under them. `grammar_structure` caught it and printed the
corrected rows, which is the reconciliation working.

Worth recording why this reached a merge at all: `--test capability` is the only
harness that runs these cells, and it was off-limits to every agent on thermal
grounds (it runs all 437 cells with container executions, no per-cell
selection). Each agent verified honestly inside the limits it was given, and
the limits had a hole exactly the shape of this bug.

@@ -239,1 +242,3 @@
-        return None;
+        return interner
+            .newtype_scalar_target(id)
+            .map(|leaf| Ty::nominal(leaf, Vec::new()));

@@ -70,0 +116,4 @@
+/// One rule otherwise, read off the declaration's own parameter list — no type
+/// name, and no per-parameter *kind* (a tag and a size are treated alike; the
+/// question is only whether the bare name fills the slot).
+fn under_determined(id: TypeId, ret: &Ty, interner: &Interner) -> bool {