http: a client — response parser, url type, https routing point

[alii]

scarlet/http was server-only. This adds the response parser, a Url type, and the routing point that picks cleartext or TLS — unblocked by #26 landing TLS.

Transport shape: parameterise, not a sum type

Socket and TlsSocket are deliberately different types, so the three shapes were: duplicate per transport, a Transport sum the client matches on, or parameterise over the transport's operations.

Duplication loses because the response reader is where framing bugs live, and two copies means two places to fix one smuggling reject.

The sum type is not absurd — it keeps the types across three small matches. It loses because it is closed, and what it forfeits is a transport backed by a Binary, which is how the whole request/response path is tested with no network. Adding a Fake arm to a production sum to recover that is a test hook on the wire.

Io was made opaque and then reverted: opaque leaves plain/secure as the only doors, which shuts out the in-memory transport that motivated the shape in the first place.

What parameterising gives up, stated rather than hidden: nothing proves all three closures address one connection. That needs an existential — Io(h) would carry it, but then Io(Socket) and Io(TlsSocket) differ and connect has no return type. Filed T-217; the cost is one unenforced invariant on a public type.

One grammar, not two

The parser is a VM op, so the header block below a status line is literally parse_header_block — the same code as a request head. That is the load-bearing claim, and plant 2 proves it: dropping the whitespace-before-colon reject turns both goldens red from one edit, http_parse and http_response together.

Plants — all watched red, all restored

plant	observed
status code defaults to 200	header bytes returned as a body: code=200 body="Content-Length: 0\r\n\r\n"
drop whitespace-before-colon reject	both goldens red from one edit
body shorter than Content-Length	body="he" instead of Err Transport(UnexpectedEof)
response framing falls back to the request rule	close-framed reply truncated to empty
drop request-header validation	X-Injected: yes on the wire as its own header
rename client.get	export check fires

One plant was inert and was replaced — disabling the obs-fold reject left the golden green, because obs-fold is caught twice. That is the only reason the check that shipped is known to be real. Two further plants tripped checks the author did not write: a no-progress read_more made read_head spin rather than fail, and adding 7 ABI slots without binding them tripped fixture_binds_every_slot.

Gates

fmt 0 · clippy 0 · gen-editor-syntax --check 0 · test --workspace 0 — 37 binaries, 1265 passed (baseline 1264; +1 is the new golden).

core_ir goldens: pure +512 type-id shift, 28 line pairs, 2 ids. The shift checker was itself planted twice — a changed variant index gives NOT A SHIFT: - ctor 7936.0 / + ctor 8448.1, an inconsistent id gives type id 7936 maps to both 8448 and 9999.

Deliberately not done

No connection pool, so T-118's own DONE-WHEN ("over a pooled connection") is not met — every request sends Connection: close and an https request is a handshake. Filed T-220, and T-165 should not be closed on this alone.

No read deadline on either scheme. net/tls has no read_within (T-128), and giving only the cleartext adapter a deadline would make liveness depend on the URL scheme — invisible until the https request is the one that hangs.

Also filed T-219: dis.rs's constant-pool ceiling was < 600, measured at 520 without these files and 604 with. Dedup was working; the ceiling was snug against stdlib size with 15% headroom. Raised to 1000, but the assertion tests the wrong thing — the failure it guards is multiplicative.

Goes stale on merge

The website's TLS page says "There is no HTTP client in the standard library. An HTTPS request is request bytes written by hand over a TlsSocket." Both sentences are false as of this branch. Separate repo, needs its own ticket.