fix(inkling): measure RAM on Windows so auto-cap is not 16 experts

[kevin9327]

Symptom

On Windows, coli serve --model <inkling> with no --cap always sizes the expert LRU to 16 experts/layer, and /health reports ram_total_gb: 0.0 / ram_avail_gb: 0.0. This host had 13 GB free while that happened. macOS already measured RAM; Windows took the #else return 0 branch.

The auto path is:

cap = avail > 0 ? (avail*0.80 - 4e9) / (slot * sparse_layers) : 16

Zero available RAM is treated as "use 16". There was no warning.

This is the Inkling half of the same class as #1601 (OLMoE Linux-only probe). It does not overlap #1611, which changes olmoe.c.

Fix

• mem_avail_bytes() calls compat_mem_available_gb() (Linux MemAvailable, macOS free+inactive+purgeable, Windows ullAvailPhys/commit). If that is still 0, print one line and keep the historical 16-slot fallback.
• serve_hwinfo() fills total/available from compat_meminfo_gb() when /proc/meminfo is absent, so /health stops reporting 0.0 GB. Windows also fills cores; Darwin fills the CPU brand.

--cap is unchanged.

Tests

tests/test_inkling_mem_probe.py:

• mem_avail_bytes must call the shared probe and must not have #else return 0
• serve_hwinfo must call compat_meminfo_gb
• on Windows, if the old branch is still present, fail with this host's free RAM (13 GB here)

tests/test_inkling_cache_index.c (CI make test-c) asserts mem_avail_bytes() is > 0 and within 0.25 GB of the shared probe.

Fail-before on unfixed dev:

AssertionError: 'compat_mem_available_gb' not found in mem_avail_bytes
AssertionError: mem_avail_bytes() returns 0 on Windows; this host has 13.2 GB free, so auto cap would stay at 16 experts/layer

Validation

• Targeted: python -m unittest tests.test_inkling_mem_probe fails on unfixed dev, passes after
• Baseline: tests.test_resource_plan tests.test_env_defaults tests.test_inkling_mem_probe -- 106 OK
• make -C c check -- local Windows host has no gcc/make; the Python pin is the gate here. test_inkling_cache_index is already in TEST_BINS.
• CUDA changes were tested with make -C c cuda-test (not applicable)
• Performance claims: none

Compatibility

• The default CPU build remains dependency-free
• No model files, generated binaries, or benchmark artifacts are included

AI-assisted (Grok)