fix(beam): read request headers from Cowboy instead of returning empty

[dbrattli]

The gap

On the BEAM target both request-header accessors were placeholders:

member x.GetTypedHeaders() : RequestHeaders = RequestHeaders(ResizeArray())  // always empty
member x.Headers = HeaderDictionary()                                        // always empty

So mustAccept and content negotiation could never resolve, and no handler could read a request header at all. An endpoint gating on Authorization had to reach past Giraffe to cowboy_req:header/3 itself. Tracked in FOLLOWUPS.md; the JS and Python backends both read their real header source.

The binding was never missing — Fable.Beam.Cowboy.CowboyReq.headers has been there all along. Giraffe just wasn't calling it.

The fix

src/beam/HttpContext.fs — both accessors read cowboy_req:headers/1 through a shared private HeaderPairs. Cowboy is already the normalising layer (lowercase names, duplicates folded into one comma-joined value), so this is a map read rather than a reimplementation of header semantics. The two-element rows it yields match the shape the Python/ASGI backend builds from scope["headers"], so mustAccept behaves identically across targets. Explicit for loops rather than Seq.map, per the Ref-vs-list caveat the file already documents.

test/beam/TestContext.fs — the other half of the gap, and the reason the suite couldn't have caught this: it accepted a ?headers argument and silently dropped it, while the JS and Python factories both wired theirs in. It now seeds them into the fake Req map under headers, lowercased on the way in to match what Cowboy hands a real handler.

test/shared/HandlerTests.fs — un-skips the three Accept-header cases on BEAM, and adds two tests covering reading an arbitrary header (Authorization, X-Request-Id) and case-insensitive lookup. Those two are the case Accept-only coverage wouldn't have pinned.

Test results

Target	Before	After
BEAM	94 passed / 7 skipped	99 passed / 4 skipped
JS	100 passed / 1 skipped	102 passed / 1 skipped
Python	101 passed	103 passed

The 4 remaining BEAM skips are the unrelated routef typed-capture divergences (%i / %O / %u), still tracked in FOLLOWUPS.md. Fantomas clean.

🤖 Generated with Claude Code