fix(planner): export K3_EXPERT_GB and GLM53_EXPERT_GB from the plan

[kevin9327]

Symptom

coli plan, coli doctor and --auto-tier size a RAM expert cache, then export RAM_GB and COLI_PLAN_CAP. Two engines do not read those as the cache size:

• Kimi K3 sizes the expert LRU from K3_EXPERT_GB (default 8 GB). RAM_GB is only a ceiling. main() never takes argv as a cap, so the gateway's COLI_PLAN_CAP is ignored.
• GLM-5.3 sizes the cache from GLM53_EXPERT_GB and never reads RAM_GB.

Unfixed dev, a Kimi plan with tens or hundreds of GB of warm experts still started with the 8 GB default. --auto-tier looked applied. The cache was not.

Root cause

environment_for_plan always set RAM_GB from the whole-process budget. That is the right variable for colibri, OLMoE, V4. It is the wrong variable for Kimi and GLM-5.3. #1585 stopped advising colibri.c-only knobs on those engines; it did not start advising the knobs they actually read.

Fix

After the colibri-only auto-tune block, export the family cache knob from expert_cache_bytes:

• kimi -> K3_EXPERT_GB
• glm53 -> GLM53_EXPERT_GB

coli plan / coli doctor print it. --auto-tier applies it (setdefault, so an explicit value still wins). coli tune then starts its Kimi resource sweep from the planned cache instead of 8 GB.

Tests

In c/tests/test_resource_plan.py:

• test_kimi_plan_exports_the_expert_cache_knob_the_engine_reads: a two-layer Kimi fixture gets K3_EXPERT_GB in tune, in format_plan, and in environment_for_plan; an explicit K3_EXPERT_GB=3.5 is kept.
• test_glm53_plan_exports_the_expert_cache_knob_the_engine_reads: the same for GLM53_EXPERT_GB, and no Kimi key leaks onto GLM-5.3.

Fail-before on unfixed dev:

AssertionError: 'K3_EXPERT_GB' not found in {}
AssertionError: 'GLM53_EXPERT_GB' not found in {}

The existing OLMoE test still requires tune == {} (no colibri knobs, and no *EXPERT_GB either: OLMoE reads RAM_GB).

Validation

• Targeted: the two new tests fail on unfixed dev, pass after
• Baseline: python -m unittest tests.test_resource_plan tests.test_doctor tests.test_autotune tests.test_env_defaults tests.test_doctor_engine_arch tests.test_v4_cli -- 200 OK
• make -C c check -- Python-only change; the modules above are the planner/doctor/tune surface
• CUDA changes were tested with make -C c cuda-test -- not applicable
• Performance claims include hardware, commands, and repeatable measurements -- none

Compatibility

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

AI-assisted (Grok)