F1: review follow-ups from the chain

[milyin]

Follow-up step of #6. Stacked on #18, at the tip of the chain.

Addresses the review findings raised against #7–#15. They land here rather
than in the branches they were raised on, so nothing in the chain has to be
rewritten and every stacked PR below stays mergeable as-is.

Correctness

Finding	PR	Fix
data class accepted a plain constructor parameter	#7	rejected in both data() and ctor_param() — Kotlin requires every primary-ctor parameter of a data class to be a val/var
value class accepted a var or plain parameter	#7	rejected — a value class wraps a single read-only property
KtCompanion::named("") rendered companion object with a dangling space	#8	panics; KtCompanion::new() is the anonymous form
a companion could not express : Base() through the builder	#9	KtCompanion::extends added
.modifier("external ") bypassed the guard	#11	each whitespace-separated word is checked, not the exact string

The companion fix is the most substantive: rendering and import collection
already supported a companion's superclass, but no builder reached it, so a
consumer had to assign supertypes.superclass directly — bypassing the
one-superclass invariant #9 exists to enforce. Both extends paths now share
KtSupertypes::set_superclass, so neither can drift.

.modifier("external inline") is also caught now — modifier strings legitimately
hold several keywords ("final override" appears in the downstream consumer),
so an exact-string compare was never sufficient. A modifier that merely
contains the substring (externalish) still passes; there is a test.

Diagnostics quality

• Scope paths no longer gain a leading / in the root package (B1: check that every declared name is a legal Kotlin identifier #15) —
  Outer, not /Outer. One scope_join helper replaces nine format! sites.
• Import-collision diagnostics report against the first-registered FQN
  (E1: diagnostics infrastructure #14) — which is the one ImportSet actually gives the simple name to. With
  three colliding imports the report previously named whichever was seen most
  recently, so it disagreed with the behaviour it was describing.

Documentation

KOTLIN_BANNER's doc said an override "falls back to" the constant, which is
backwards — the fallback applies when there is no override (#13).

One comment not acted on

#7 flagged that matches!(c.kind, KtClassKind::Value { .. }) "attempts to move
kind out of &KtClass, which does not compile". It does compile: the pattern
binds nothing, so nothing is moved. The code has been building and passing tests
since that commit.

Also already resolved by the chain

#14 flagged that check_top_level_names ignored KtDecl::FunInterface. Correct
at that commit — E1 was a deliberate verbatim move — and fixed two steps later
by B2 (#16), which replaced that function with the namespace-aware walker.

Verification

All 88 pre-existing tests pass. Thirteen added, covering every fix plus the
must-not-fire cases (val+var data properties accepted, externalish accepted).
101 tests, 10 doctests, cargo clippy --all-targets clean.