gate: don't strand the terrain generator; bound the splash's waits

[dsrw]

Loading a level from inside another one (clicking a sign link that evals load_level("welcome")) held the splash for the full 30s spawn-gate timeout, then revealed with spawn gate: deadline exceeded. The level had finished loading well before that — the process was completely idle for the whole wait (snapshots_delta=0, no meshing), with build_water the only unit still reporting pending work.

Root cause

build_node drops the terrain's generator and stream on RESETTING.added and restores them on RESETTING.removed. But Build.reset() adds and removes RESETTING inside one worker tick, and a build's node is created in the middle of that window — its script runs right after the thing joins state.things. The node can see the added, as its watch replaying a flag that is already set, while the removed that preceded its registration never arrives as an edge.

The terrain is then left with no generator and no stream at all, which turns streaming off wholesale in VoxelTerrain::_process (stream_enabled false): no blocks are ever requested, has_stream_started() never becomes true, and the unit publishes the not-started floor of 1 forever with no work behind it. The spawn gate waits on that until it times out.

keep_generator_in_sync now derives the generator from the flag each frame instead of trusting its edges.

Evidence: booting tutorial-3 stranded three units deterministically — build_3npxfmddql9kx, build_mlsxf11t2oaz6, build_wevpat48mxjhe each logged "stream never started" on both pre-fix runs. After the fix, none, across every run since.

Bounding the gate

So no single unit can hold the splash again:

• The not-started floor expires after 3s. It exists to stop a terrain reading "done" before its required set is computed — a frame or two, measured well under a second even for the 6528-chunk terrain. Past that nothing is coming, and it logs the unit it dropped plus whether it still had a generator.
• The warm-up lifts after 1s even if get_frames_drawn() never advances. It waits for two real draws behind the splash, which needs the engine to actually be drawing; a minimized or occluded window (every MCP-managed and test instance) freezes that counter, and the splash never lifted at all, with nothing logged. This was a second, independent way to hang the splash forever.
• Settle timeout 30s → 10s, and the deadline error names the blocking units with their counts (blocking=["build_water=1"]) instead of just the whole gate set.

Testing

nim test_all passes. Switch timings for tutorial-3 → welcome: 628ms with a drawing window, 1.58s minimized (previously: 30s timeout, or never).

🤖 Generated with Claude Code