MEAS-001: granular resource receipts (RSS/VRAM/CPU/IO/faults/energy) for bench, CRUX, dogfood, CI ledger #4522

Operator (2026-09-27): our instrumentation must show not only timing but granular resource usage.

Audit (cop, 2026-09-27): today every receipt we have is timing-only.

• apr bench --json (crates/apr-cli/src/commands/bench.rs:356-395) emits no resource fields.
• Build ledger peak_rss_mb is null in 1084 of 1095 receipts, and free_disk_gb is null too.
• The serve-bench MemoryResults/EnergyResults hold estimates or constants.
• CRUX D-03 --report-peak-memory is a draft.
• CiAssertions.max_memory_mb is dead code.
• The dogfood C1 RSS check only WARNs.
• No resource number is gated.

Rows. Every row is measured, not estimated, and every field names its source:

• R1 (aprender-52): apr bench/qa/profile JSON gains resources{}:
  • peak_rss_bytes (getrusage ru_maxrss)
  • cpu_user_s and cpu_sys_s
  • minor and major page faults
  • voluntary and involuntary ctx switches
  • threads_peak
  • io read and write bytes (/proc/self/io)
  • vram_peak_bytes (NVML/cuMemGetInfo high-water, sampled)
  • energy_j (RAPL, when readable; else null with a reason)
    Also: close CRUX D-03 and replace the serve-bench estimates with real readings.
• R2 (aprender-5d): gates.
  • Wire CiAssertions.max_memory_mb plus a VRAM ceiling into perf-matrix.yaml, with contract thresholds.
  • Turn dogfood C1 RSS from WARN into FAIL.
  • Case table: a planted over-budget run must go RED.
• R3 (infra-33): CI build ledger reads on-runner cgroup v2 (memory.peak, cpu.stat, io.stat, memory.events) per job, so no field stays null. The arbiter logger's cpu_seconds must match the spec.

Done when a real receipt from nightly and from an rc shows every field non-null (or null with a stated reason), and one gate has been proven RED by mutation.