chore: release v0.0.4

CHANGELOG.md

@@ -7,6 +7,80 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.0.4](https://github.com/KarpelesLab/rustygo/compare/v0.0.3...v0.0.4) - 2026-10-07
+
+### Other
+
+- CI has been red since October 4th, and this is why
+- The M:N scheduler, landing at one worker
+- What more than one worker gets wrong is memory, not a context
+- The parallel program runs at the GOMAXPROCS it is given
+- The roadmap says what more than one worker still gets wrong
+- reflect's TryRecv had the race select had
+- One worker by default, and what is known about why
+- A collector that loses a race was still holding its roots on its thread
+- Two workers by default, until three stop corrupting a context
+- Which goroutine is leaving is the scheduler's to say, not the caller's
+- Stopping the world, and GOMAXPROCS at the number of CPUs
+- Goroutines on worker threads, and a park that cannot lose a wakeup
+- Runtime state that belongs to a thread, and state that belongs to the program
+- 74 programs, and the whole set passes under torture as well as plainly
+- reflect.ArrayOf, answered from the arrays a program already has
+- A struct holding a func value does not have gc's layout
+- The standard descriptors, closed by their own finalizers
+- 127 of Go's own tests, with recover.go whole
+- recover sees through the frames that stand in for a deferred call
+- The missing descriptor that is not missing
+- A timer that must fire while a goroutine spins, not only while it parks
+- HTTP/2, and what the BoGo suite says
+- What crypto's own tests say, and the two things crypto/tls still trips on
+- HTTP/2 over TLS, and large bodies with it
+- The clock syscall reads directly, and the bridge into libc it never crosses
+- reflect.MakeFunc, which is Value.Call run backwards
+- A named map or channel that contains itself, through the same name
+- A pointer the collector cannot see, because it became an integer
+- The metrics and the GC statistics the collector can actually answer
+- Why an allocation test fails, now that it has been measured
+- gRPC, protobuf, Prometheus and pgx compile, so the table should say so
+- The three tables, measured on this tree instead of three trees ago
+- What the threaded scheduler found before it found its own bug
+- A map lookup keyed on bytes does not need the string either
+- One copy where string and []byte conversions made two
+- The HTTPS program makes its own certificate
+- The runtime hooks that each stopped one package building
+- A named func type that returns itself
+- one document for the packages a sweep had time for and the rest
+- `noasm`, which is what libraries outside the standard library call `purego`
+- which failures to count apart is the report's judgement, not the sweep's
+- What the front page claims, brought back to what is true
+- One conversion, nine comparisons, and none of them made a string
+- Where the allocations rustygo makes and gc does not actually are
+- How a program outside the standard library reaches the kernel
+- Comparing bytes as a string without making the string
+- The types reflect can be asked to make, and the one it can make
+- Channels through reflection, and what a type says about ranging
+- What the compiler can express, measured for the whole library at once
+- measure what cannot be built, and in what order to fix it
+- Sleep gives up the processor however short it is
+- A read deadline is a moment, and it was being read as a duration
+- A differential program for the file layer
+- $GODEBUG reaches the standard library, and SIGSYS can be ignored
+- A tick says when it was due, a reset does not queue a timer twice, and a
+- A goroutine that only makes system calls lets the others run
+- A loop that does not wait for rustc
+- The same system call on every thread, where there is one
+- Signals reach Go, and a broken standard output ends the program
+- What os, time and net/url could not even type-check
+- The FIPS module's XOR, under Rust's rules about alignment
+- A big array is filled in place, not built on the stack first
+- HTTPS, end to end, in one process
+- A run cap that a TLS handshake fits inside under GC torture
+- encoding/json passes its own test suite
+- A named type that contains itself
+- Do not commit the worktrees agents work in
+- Build one package at a time
+- Goexit through a frame that defers, and a test binary's own directory
+
 ## [0.0.3](https://github.com/KarpelesLab/rustygo/compare/v0.0.2...v0.0.3) - 2026-09-29
 
 ### Other

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "rustygo"
-version = "0.0.3"
+version = "0.0.4"
 edition = "2024"
 rust-version = "1.89"
 description = "Runtime for Go packages compiled to Rust by the rustygo compiler"