JangPress RSS accounting after expert deactivation on macOS

Facts

• `TASK_VM_INFO.resident_size` is read from the `phys_mem` ledger, and `external` and `internal` are separate ledgers read next to it. The `phys_mem` ledger is the pmap-resident total, so it includes clean file-backed pages that a process has faulted in. [source]
• `phys_footprint` is documented in the kernel as the sum of (internal - alternate_accounting), (internal_compressed - alternate_accounting_compressed), iokit_mapped, purgeable_nonvolatile, purgeable_nonvolatile_compressed and page_table. It has no term for `external` (file-backed) pages. [source]
• Consequence for a mmap safetensors bundle: clean expert pages that stay mapped appear in `resident_size` (RSS) and never in `phys_footprint`. A tool that reports footprint shows a small idle number for a mapped bundle even if no page was ever dropped. [source]
• Consequence for `MADV_DONTNEED`: the call leaves pmap entries in place (existing dossier), so `resident_size` should not fall at the call; it falls only when the pageout scan steals the deactivated pages. `phys_footprint` is unaffected either way. [source]
• `iokit_mapped` counts the whole size of every IOKit mapping in the task "regardless of clean/dirty or internal/external state", and `vm_map_iokit_mapped_region` credits both `iokit_mapped` and `phys_footprint` with the full byte count at map time. [source]
• In the page-query code, non-purgeable IOKit internal memory is reported as present and dirty for its entire virtual range ("phys_footprint includes the entire virtual mapping"). [source]
• No dated history for JangPress accounting claims was found beyond the 2026-05-04 guide date already in jangpress-cold-expert-eviction-for-moe-larger-th.md. [source]
• If a runtime wraps mmap pages as no-copy Metal buffers and those buffers are counted as IOKit mappings, `phys_footprint` would be charged the full mapped size at map time. That would explain a footprint near 130 GB at first prefill, but no source states how vMLX's no-copy buffers are accounted. [source]
• The Kimi runtime README reports "~130 GB process footprint" before token 1 and "about 1 GB" idle RSS for the same bundle. The two numbers can both be correct only if they come from different ledgers or from different moments; the README does not name its tool. [source]
• Activity Monitor, `ps` and `footprint` read different ledgers, so a before-and-after comparison of JangPress must use one tool throughout. [source]
• Reading of "idle RSS about 1 GB". Reading A: the kernel dropped the cold experts after `MADV_DONTNEED`. Reading B: the reported number is a footprint-style figure that never counted clean mapped file pages. The kernel source supports B as sufficient and A only under reclaim pressure; no source measures both ledgers on a JangPress run. [source]
• Which tool and ledger produced the jangq "about 1 GB" and "130 GB" figures? [source]
• Whether vmlx-swift-lm's no-copy `MTLBuffer` over mmap pages is accounted as `external`, `iokit_mapped` or neither. [source]
• A direct test: run a JangPress load and read `TASK_VM_INFO` (`resident_size`, `external`, `phys_footprint`) plus `mincore` before and after a DONTNEED sweep. No such test is published. [source]
• `TASK_VM_INFO.resident_size` is read from the `phys_mem` task ledger, with separate `external`, `internal`, `internal_compressed` and `reusable` ledgers returned alongside it. [source]
• The kernel comment defines `phys_footprint` as (internal - alternate_accounting) + (internal_compressed - alternate_accounting_compressed) + iokit_mapped + purgeable_nonvolatile + purgeable_nonvolatile_compressed + page_table. [source]
• The ledger comment defines `internal` as the task's anonymous memory and `internal_compressed` as internal memory held by the compressor, "no longer actually resident for the task". [source]
• `iokit_mapped` is the total size of all IOKit mappings in the task regardless of clean, dirty, internal or external state. [source]
• `vm_map_iokit_mapped_region` adds the mapped byte count to both the `iokit_mapped` and `phys_footprint` ledgers. [source]
• Because `phys_footprint` has no `external` term, clean file-backed mapped pages count toward `resident_size` but not toward `phys_footprint`. [source]
• A `MADV_DONTNEED` sweep changes neither ledger at call time; `resident_size` falls only after the pageout scan frees the deactivated pages. [source]
• The jangq "about 1 GB idle RSS" and "about 130 GB footprint" figures are not tied to a named ledger, so they cannot be compared with each other. [source]