tapped's gate is an arity error: the last rest tuple

[btravers]

@btravstack/testing's tapped is the last compile-time gate in this repo that
rides a conditional rest tuple — the shape #66 removed from start precisely
because an arity error never prints a type. #82 documented it honestly rather
than fixing it; this is the fix.

What it prints today

packages/testing/src/tapped.ts:34-36:

...gate: [Exclude<InstanceType<P[number]>, X>] extends [never]
  ? []
  : [error: "NOT EXPORTED", missing: Exclude<InstanceType<P[number]>, X>]

Tapping a port the module does not export (measured, scratch file in
packages/testing/src, npx tsc -p tsconfig.json --noEmit):

error TS2554: Expected 4 arguments, but got 2.

That is the whole message. Neither NOT EXPORTED nor the offending port
reaches the reader. Both are in the parameter's type and answer one slot at a
time, which is the technique /reference/di/entry-points#the-gate teaches:

error TS2345: Argument of type '0' is not assignable to parameter of type '"NOT EXPORTED"'.
error TS2345: Argument of type 'number' is not assignable to parameter of type 'Secret'.

Secret is the port — reachable in two hand-spelled steps, printed in none.

The fix, and the one thing that has to be measured

start moved its marker onto the module parameter and the sentence started
printing, because a parameter type is something TypeScript prints and an arity
count is not. The same move applies here, with one open question: which
parameter carries it.

• ports reads best — the mistake is the ports array, and a sentence like
  "NOT EXPORTED — the module exports no port named Secret" lands on the
  argument that is wrong.
• module is the closer analogue to start, and the safer inference bet.

The risk is stated in packages/core/src/start.ts:139: a conditional type in an
inference-bearing position can defer that parameter's inference and collapse a
type parameter to unknown. For start it was measured not to — X still
infers from the Module<X, …> half of the intersection. Here P is
const P extends readonly AnyPort[], inferred from ports, and a const type
parameter under an intersection is not obviously the same case. Measure both
placements before choosing; do not reason it out.

Whether the port's name can reach the sentence is a second measurement.
@btravstack/http's UNDECLARED KEY gets its key in via a template literal
because its mapped type is keyed by K (#82, ac12553). tapped has an
Exclude<...> of port instances rather than a key, so the same trick may not
be available — if it is not, "NOT EXPORTED — this module exports no such port"
printed in full is still strictly better than an arity count.

What moves with it

• packages/testing/src/tapped.ts — the gate, and its TSDoc, which currently
  says the gate "refuses a port module does not export, at this call site".
• docs/reference/testing.md — the #the-tap-gate-an-arity-error section, in
  full: its anchor is its claim, so the anchor changes too, and every inbound
  link with it (docs/how-to/test-an-application.md is one).
• docs/explanation/compile-time-wiring.md — "Three gate mechanisms live in
  this repo that a composing application meets… A fourth lives in the test
  harness". If tapped stops being a rest tuple, the fourth stops being an
  arity gate and that whole passage is restated.
• A *.test-d.ts pinning the new arm. packages/testing has no type tests and
  no tsconfig.test-d.json today — packages/temporal was in exactly this
  position before Keyed dependency records for Provider — positional deps are the highest-frequency DX wart #65 and gained both, so there is a worked precedent.
• A changeset. @btravstack/testing only; the nine move as one anyway.

Not in scope

di's own gate on Module.scoped / build / forkScope stays a rest tuple, and
that is a decision, not an oversight — docs/explanation/compile-time-wiring.md
argues it: the arity form points the blame at the call rather than deep inside a
generic instantiation, and di's missing: N is a set of ports a reader can
read off the signature, where the kernel's is one of three fixed sentences.
tapped's is a single sentence plus one port, which is the kernel's shape, not
di's.

---

Found by the whole-branch review on #82, which caught that #66's sweep had
stopped at three mechanisms when there were four.