[lucidprogrammer]

scripts/ci-local.py runs the pull-request workflow on a build host with each job boxed like a hosted runner, and starts a step only while the boxes' shared memory has room. A step marked itself as running before it waited for that room, so two steps waiting at once each saw the other as running and both waited forever; the same held for two jobs waiting to rerun after the host ran out of memory.

What changes

• A step takes its running unit only when it is admitted, under the runner's lock, so a waiting step never counts as a running one.
• A job waiting to rerun counts only boxes that are running a step, not other boxes that are also waiting.

Verification

• A stub run of two waiting steps with no free memory: before the change both wait forever; after it, one starts and the other starts once the first finishes.
• GitHub Actions (Linux) on 25e6b3e: https://github.com/context-graph-ai/context-graph/actions/runs/42762416150 — green, 13 jobs passed, 1 skipped. The first attempt failed only on the intermittent concurrent observation test that PR tests: accept a duplicate-key refusal from the losing writer in two concurrent observation tests #10 fixes; the rerun of that job passed.
• Reviewed by a second model family (SHIP).