fix: serialize frame writes per socket, not per writer

[scgopi]

Fixes the "unrecognised command — graphcoded may be older than the client" error seen on first open of a new workspace in 0.1.46-beta1, which a quit and relaunch clears.

It isn't version skew

Nothing is out of date — the bytes are spliced. A frame goes out as two writes (a 4-byte length header, then the body), and nothing gave a connection a single writer:

Direction	Concurrent writers on one socket
app → daemon	every DaemonConnection.send hops onto DispatchQueue.global(); AppFeature.start() merges three at launch — listRecentProjects, restoreOpenProjects, openGlobalGraph
daemon → app	ProjectRegistry and GraphStore are separate actors, plus the skew reply in graphcoded/main.swift:136

Overlap two and the reader consumes header A followed by body B, then reads a length belonging to someone else's message and is wrong from that byte onward — it either decodes garbage (which raises exactly this error, the only place that string exists) or blocks waiting for bytes that never arrive.

Why a new workspace specifically: the three launch commands wait together on connectWithBackoff while the freshly bootstrapped daemon comes up, then resume at the same instant — the widest possible collision window. On the next launch the daemon is already listening, the sends stagger, and it doesn't land. That is precisely "first time only, restart is fine".

The fix

FramedMessageIO.writeFrame now takes a lock for the descriptor it is writing to, so a frame can never interleave with another on the same connection. One call site changed, all five writers fixed.

Per descriptor rather than one lock for the process: a write blocks while its socket buffer is full and the daemon broadcasts to every connected client, so a single lock would let one wedged client stall the writes to all the others. Locks are kept rather than reclaimed on close — descriptor numbers are small and reused, so the table stays about as big as the peak connection count, and a reused number gets the same lock back.

Verification

concurrentWritersDoNotInterleaveTheirFrames — 8 writers, 64 KB bodies (larger than a socket buffer, so a single write cannot complete in one syscall and the interleaving is certain rather than occasional):

	Result
Before	❌ 1 of 8 frames arrived; reader then hung on a spliced header until the 30s timeout
After	✅ passes in 0.005s

Full suite: 989 tests passing. swift format lint --strict clean.

🤖 Generated with Claude Code