iili: model HEAD and delete WithHeadAsGet once smithy-cpp frames HEAD responses

[aaylward]

Follow-up to #1433, closed by #1435 with a service-local workaround rather than the modeled contract.

Where it stands

HEAD /iili/v1/r/{slug} answers 302 with Location today, via WithHeadAsGet in domains/iili/apis/iili/smithy_handler.cc: it re-dispatches HEAD as the GET it mirrors and clears the body on the way out. main.cc and both test harnesses share that composition, and wire_test.cc pins the 302 and the 404.

It is deliberately not modeled. Routing was never the obstacle — smithy-cpp's router buckets by method and only checks conflicts within a bucket, so @http(method: "HEAD", uri: "/iili/v1/r/{slug}") would register cleanly beside the GET. The obstacle is framing: ToWireResponse copies the body and calls prepare_payload() without ever seeing the request method, so a modeled HEAD operation would put the generated {} body on a HEAD response — forbidden by RFC 9110 §9.3.2, and on a keep-alive connection the client reads those two bytes as the start of the next response. The handler cannot suppress it: the body comes from the generated serializer and the transport owns framing.

Filed upstream as muchq/smithy-cpp#192.

What to do here, once the pin picks that up

1. Add the HEAD binding for Redirect to domains/iili/apis/iili/model/iili.smithy.
2. Delete WithHeadAsGet and its declaration, and unwrap the three call sites (main.cc, wire_test.cc, production_chain_test.cc).
3. Keep both HEAD wire tests exactly as they are — they assert behavior, not mechanism, so they should stay green across the swap. That is the check that the modeled version is a true replacement.

Why it is worth doing

The contract is invisible where a reader looks for it. Nothing in iili.smithy says HEAD is answered, so the model understates what the service serves, and a generated client has no HEAD operation. The workaround also lives only in this service; the modeled form would come from the same place every other route does.

Not urgent — behavior is correct on the wire today. This is about where the truth is written down.

[aaylward]

Done in #1445, merged as 5fd7ce7. It did not autoclose because the PR said Closes #1433 — the already-closed predecessor — and never referenced this issue. My mistake: I found WithHeadAsGet by grep and reconstructed the rationale this issue already had.

All three steps landed, with one departure from step 3.

1. HEAD binding. Added as its own operation, RedirectHead, rather than a second binding on Redirect — Smithy models one @http trait per operation. Same URI, same output shape; Redirect and RedirectHead answer from one private Location(slug).

2. WithHeadAsGet deleted, all three call sites unwrapped.

3. The two HEAD wire tests could not stay green, and that was the finding. They ran on smithy::http::Loopback, which hands back the handler's response and does no framing — it cannot observe Content-Length at all. EXPECT_EQ(response.body, "") held only because the middleware cleared the body inside the handler chain. With the middleware gone, the serializer's body reaches Loopback untouched and both tests failed.

So they were asserting mechanism, not behavior: the property they pinned was "something in the handler chain emptied this", which is exactly the thing being deleted. They could never have survived the swap.

Coverage is now split by what each harness can see:

test	harness	asserts
HeadRoutesAndResolvesExactlyAsGetDoes	Loopback	status, Location and body match the GET
HeadOnAnUnknownSlugIsTheSame404AsGet	Loopback	the 404 is the GET's 404
HeadRedirectCarriesTheGetsLengthAndNoBodyOnTheWire	Beast + raw socket	empty body, Location, declared length equals the GET's
HeadOnAnUnknownSlugIsFramedLikeItsGetToo	Beast + raw socket	the same, on the 404

The wire tests read raw bytes rather than using SocketHttpClient, which now stops after a HEAD's headers and would report an empty body whether or not the server sent one. Both fail against an un-fixed transport.

Also worth recording against this issue's premise: the old workaround was not merely unmodeled, it was wrong on the wire. Clearing the body does not omit it — the transport sizes the response from what it is handed, so HEAD /iili/v1/r/{slug} shipped Content-Length: 0 where the GET reported 2. "Behavior is correct on the wire today" held for status and Location, not for the length.

Upstream muchq/smithy-cpp#192 is merged; the pin moved 46bcd4d → a01f0d2, which also raised abseil and forced a deprecation fix in encoding.cc.

---

Generated by Claude Code