fix: at macOS memory warning, start jobs while memory fits

[SpoBo]

Closes #22.

Problem

macOS keeps its memory warning level for a long time, also with half of memory free. A running dev stack held it there, and taskguard started nothing new at warning. The browser proof that needed the stack waited behind it: 25 minutes for one proof, an hour of retries for another.

Change

At warning (macOS only), a job starts when:

• memory fits without any leeway (no noise_mem, no outside_admit, also when only long-lived jobs run), and
• no running first run still grows.

A job with a history has its whole need booked from its first second, so the next job may start at once. A first run books a guess, so new jobs wait until it has shown what it takes (the existing settling rule, at most two minutes).

Critical still stops every new job. warn_admit = false keeps the old rule. Linux has no warning level and keeps its PSI rule.

Evidence

• Unit test at_warning_jobs_start_while_memory_fits_and_no_first_run_grows.
• Simulator (src/sim.rs), with the warning level held from 40% memory: the old rule never finished 4 of 6 scenarios. The new rule finished them in the same time and with the same peak memory as with no warning at all.
• Local history (17 days): pressure was the main blocker for 16.2 job-hours of waiting. In the last 24 h, warning was on 10.5% of the time, at a median of 75% memory used.

Mixed versions

No change to queue::Entry, machine::MachineSample, the database, or job keys. warn_admit is a new setting; older versions ignore it with a warning and keep stopping at warning. Blocker::Pressure gains optional output fields only.

🤖 Generated with Claude Code