Record the 24 GB Mac's measured RAM gate and the replacement-Mac pickup - #195

What this records

The macOS host in service is RAM-limited, measured on macos-m5pro-20260924 (Apple M5 Pro, 24 GB) on 2026-09-24:

• Current load already exceeds physical memory. Excluding the gitleaks runaway, processes held 26.84 GiB of footprint plus 3.07 GiB wired on 24 GiB. There were 357.8 GiB of swap writes in 4.81 days, and two Jetsam events (09:45 and 09:46 EDT) killed 405 and then 4,255 processes.
• The RAG embedder can't stay resident. Loading qwen3-embedding:4b (4.09 GiB) leaves 2,079 MB unused with Chrome closed; the shared service already unloads it after 30 s.
• The 27B local model can't load. The Metal working-set limit is 17.76 GiB, below the 17.67 GiB of 27B Q4_K_M weights plus any cache.

Files:

• evidence/artifacts/host-upgrade-20260924/upgrade-evidence.json: the RAM gate, the replacement recommendation and the Mac role. It supersedes the projected-64 GB MacBook item of host-upgrade-20260923.
  • Recommendation: 64 GB, M5 Pro, US$4,199. Alternatives: 48 GB ($3,799), Max 64 GB ($5,099), Max 128 GB ($6,699, not justified). Apple Store prices read 2026-09-24.
  • Mac role: cockpit and coordinator. Heavy and always-on lanes run on the WSL2 workstation, reached through Claude desktop SSH sessions and Remote Control.
• evidence/artifacts/host-upgrade-20260924/foundation-alignment.json:
  • memory, RAG and all 20 foundation layers as they stand on the Mac;
  • the saturation status (0 of 32 layers closed) and the model candidates with Hugging Face metadata;
  • the user's 2026-09-24 license decision;
  • the open items for the replacement Mac.
• docs/next-host-stages.md: the macOS role row, a measured upgrade row, and replacement-Mac pickup steps.
• manifests/evidence.json: the two registrations and the updated doc hash.

Scope and ownership

• No Mac global state was changed. Its single writer is foundation-lane-setup (user decision, seathatflowsinourveins/agent-ecosystem#28); the measurement claim is posted there.
• No overlap with other open PRs. It avoids the files of Record Mac (macos-m5pro-20260924) foundation use receipts: rtk, worktrunk, gitleaks #185 (this Mac's profile and receipts), Adoption hygiene: native Claude Code pin is a floor (2.1.281), ai-memory assistant capture opt-in, ignore Claude Code local state #159 and Linux/WSL2 bootstrap keeps a newer native Claude Code: the pin is a floor (stacked on #159) #177 (the macOS platform page).
• The gitleaks runaway is bounded separately in Bound gitleaks scans on macOS: per-user lock and 6 GiB footprint watchdog (Jetsam 2026-09-24) #193.

Verification (python3.13, local)

• CI validation steps: all passed, including validate.py (5,282 hashed files), host_receipts.py validate, evidence_manifest.py --check, catalog and foundation validation, verdict, matrix, grand-list, crosswalk, wave-ledger and ecosystem checks.
• Unit tests: python3.13 -m unittest ran 4,002 tests. Inside the Seatbelt sandbox, 35 tests errored with PermissionError: Operation not permitted on local sockets. Re-running those modules outside the sandbox, every one passed except the known macOS-only test_install_path_with_shell_syntax_is_data_in_the_wrapper, which Make the llama wrapper quoting test pass on SIP-enabled Macs (wrapper unaffected) #178 fixes.

Limits

• One incident window and a few snapshots, not multi-day monitoring. Estimates are labelled in the records.
• Facts reported by the owner sessions are marked reported_by_owner.
• No candidate model was downloaded. The RAM-fit matrix is an open item, pending the user's approval of the downloads.

🤖 Generated with Claude Code