fix!: upgrade to Fable 5.11 and fix chardata handling in string bindings

[dbrattli]

Upgrades to Fable 5.11 and fixes the issues that upgrade surfaced.

Requires Fable ≥ 5.11.0, the first release with the two BEAM fixes this depends on
(#4775, #4776):
Assert.AreEqual/NotEqual now raise on the BEAM, and erased types (Pid/Atom/Shutdown) no
longer emit reflection calls that fail to resolve. Verified against the stock tool.

What this fixes

Assert.AreEqual now raises on the BEAM. Fable's BEAM backend previously lowered it to an
equality expression whose boolean result was discarded, so assertions couldn't fail. Fixed upstream
in #4775; the same gap affected Fable's own BEAM test suite. With assertions working, this repo's
suite surfaced 14 real failures, all fixed here — it's now 367/367, and it can demonstrably fail
(it reported 14, then 6, then 0 as these landed).

Chardata handling in the string bindings (8 fixes). The OTP string module returns chardata
— an iolist, or a charlist of codepoints — not a binary. F# strings are binaries, so
string:pad("hi", 5) was [<<"hi">>,32,32,32], which compares unequal to <<"hi ">>. Affects
replace_all, replace_first, the three pad arities, reverse, to_graphemes and
compose_query. Now flattened at the binding boundary with unicode:characters_to_binary/1, chosen
over iolist_to_binary/1 because it also handles charlists and codepoints above 255.

Worth knowing for anyone who has been using these: an iolist is still valid chardata, so these values
behaved correctly when passed straight to io:format, file:write or a Cowboy body — OTP flattens
them. The mismatch only bites when the value is compared, pattern-matched, or stored as a binary.

Six test corrections. The bindings they covered were all correct; each was reproduced by hand in
erl before being touched. longest_common_prefix genuinely returns 3 for that input. The receive
test sent {ping} where a nullary DU case compiles to the bare atom ping. The four jsx labels
tests used jsx:is_json/2, which doesn't accept the labels decoder option and so could never
observe it — they now decode and assert on the resulting key type.

Breaking change

string:reverse and every string:pad arity return chardata, so they can't be bound through
[<ImportAll>] — there's nowhere to convert the result. They move to the module's typed-API section,
which is where String.fs already says "functions with non-trivial Erlang return values" belong.
Module-level F# functions can't overload, so the arities become distinct names:

Was	Now
str.reverse s	reverse s
str.pad (s, n)	pad s n
str.pad (s, n, dir)	padDir s n dir
str.pad (s, n, dir, c)	padWith s n dir c

Pre-1.0 (-rc.x), and synapse doesn't use any of them, so no known downstream breakage.

Test-runner hardening

Fable 5.8+ qualifies generated module names with the OTP app name (test_maps →
fable_beam_test_test_maps), which broke test_runner.erl's test_-prefix discovery. It now
discovers tests by exports (any module exporting test_*/0), so it's robust to naming changes,
and it refuses to report success on an empty suite — a discovery bug can't look like a green
build.

Scriptorium spike

spike/scriptorium (just spike) runs Scriptorium — Nib
assertions + the Quill runner — on the BEAM against real bindings. No test_runner.erl and no
[<Fact>]: Quill's runner is the [<EntryPoint>], which Fable emits as main:main/1 and which
halts the VM with the suite's exit code. Failures come with a diff and a clickable source link
(CallerFilePath survives the BEAM).

The case for adopting it: Nib's assertions raise from F# itself rather than depending on the backend
lowering Assert.AreEqual, and Quill registers tests explicitly rather than rediscovering them by
naming convention — so neither of the two problems above is expressible. Migrating test/ off the
homegrown runner is the natural follow-up, deliberately not in this PR.

Verification

just test    # 367 passed, 0 failed
just spike   # 11 passed, exit 0

🤖 Generated with Claude Code