Replaces the response-directory fallback with what the frontend already
does for ephemeral sends: the engine hands the body back, because it is
the only copy. Guessing at a file named for an id was the store reaching
around its own abstraction, and it is exactly what must not survive the
move to blob storage.
The send reply carries the bytes when the engine returned them, and the
runtime holds them for the rest of the call that sent them, so
ctx.httpResponse.body() answers for a saved and an unsaved response the
same way. Unsaved ones now report a real contentType too, taken from the
response the send already handed back.
A send with no request behind it — a plugin's ad-hoc ctx.httpRequest.send,
GraphQL introspection — gets a generated id and a body file, but no row.
Resolving purely through the database refused those, which would have
broken auth-oauth2 the moment it moved off readFileSync, since every
request it sends is ad-hoc.
The store now falls back to the response directory when there is no row,
accepting only ids shaped the way the engine generates them. Such a
response has no stored headers, so contentType is null and text()
decodes as UTF-8 — which is what the filesystem readers did anyway.
Plugins read bodies by response id through ctx.httpResponse.body()
instead of opening HttpResponse.bodyPath themselves. The accessors are
named after fetch's, minus the single-use semantics, since the bytes are
durable and re-reading should work.
Underneath is a chunked pull over the existing plugin protocol, so the
host can move bodies off the filesystem without plugins noticing.
text() now decodes with the response's charset rather than assuming
UTF-8, and the buffering accessors refuse past 32 MiB and point at
chunks().
Gregory Schier