fix(planner): auto-tune advised colibri.c-only knobs for every engine

[kevin9327]

Symptom

coli plan and coli doctor print an auto-tune: block on every model family except GLM-5.3. On all of them except GLM-5.2, each knob in that block is one the engine never reads. --auto-tier exports the same knobs into the engine's environment.

Unfixed dev, Windows, tiny fixtures:

$ coli plan --model qwen36_tiny_c
  ...
  auto-tune:
    DRAFT=0            compute-bound: MTP batch overhead exceeds yield
    COLI_CUDA_PIPE=1            single GPU: S=1 pipeline gate

$ coli plan --model dsv41_tiny          # DeepSeek V4.1 Flash
  ...
  auto-tune:
    DRAFT=0            compute-bound: MTP batch overhead exceeds yield
    PIN_GB=all          enough RAM for full expert residency

A user following that advice sets variables that change nothing. coli tune also starts from those values as its baseline environment.

Root cause

_auto_tune can emit DRAFT, PIPE, COLI_CUDA_PIPE, COLI_NUMA (+ the numactl hint) and PIN_GB. git grep '"<name>"' -- '*.c' '*.h' '*.cu' '*.inc' '*.mm' finds each of the five in c/colibri.c and nowhere else.

58ed660 already hit this for glm53 and excluded it: "glm53 ... does not consume the generic DRAFT/PIPE/PIN/NUMA knobs below. Recommending them is worse than leaving them unset because coli tune then reports changes the engine ignores." The same holds for inkling, kimi_k3, olmoe, qwen36, qwen38, deepseek_v4 and deepseek_v41, and they were not excluded.

Fix

Return no knobs for any engine_group other than colibri-core, the group GLM-5.2 uses. This extends the glm53 exclusion to the other families. GLM-5.2 is unchanged. The engine_group=None default, which only unit tests use, keeps the generic behaviour.

Tests

In c/tests/test_resource_plan.py:

• test_only_the_colibri_engine_gets_its_own_knobs runs four bottleneck and GPU shapes (disk, resident CPU, one GPU, mixed with two GPUs, on one or two sockets). With colibri-core they emit all five knobs and the hint. Every other engine_group in the registry (read from all_families(), so a new family is covered automatically) gets {}.
• test_sibling_plan_advises_no_colibri_knob builds a plan for a two-layer OLMoE container and for the GLM fixture already in the test class, with the same resources:
  • GLM still gets DRAFT and PIN_GB.
  • OLMoE gets an empty tune, no auto-tune: block in format_plan, and none of the five keys in environment_for_plan.

Fail-before on unfixed dev:

FAIL: test_only_the_colibri_engine_gets_its_own_knobs (...) (engine_group='deepseek_v4', case=('disk', 0.5))
AssertionError: {'DRAFT': {'value': '0', 'reason': 'low hi[312 chars]sts'} != {}
...
FAIL: test_sibling_plan_advises_no_colibri_knob (...)
- {'DRAFT': {'reason': 'compute-bound: MTP batch overhead exceeds yield',
-            'value': '0'},
-  'PIN_GB': {'reason': 'enough RAM for full expert residency', 'value': 'all'}}
FAILED (failures=29)

That is 7 engine groups × 4 shapes, plus the plan test. The glm53 subtests already passed.

Verification

• Every test module that touches the planner or its output passes: test_resource_plan, test_doctor, test_cli_output, test_env_defaults, test_openai_server, test_registry_engine_agreement, test_analysis_cache, test_v4_cli, test_autotune, test_cluster_sharding and test_efficiency_report. That is 573 tests, OK (3 skipped), with PYTHONUTF8=1.
• End to end, Windows, on tiny fixtures built with the repo's generators (make_qwen36_tiny.py + convert_qwen36.py, make_dsv41_tiny.py, make_olmoe_tiny.py + convert_olmoe_merged.py, make_glm_oracle.py):
  • coli plan on qwen36_tiny_c, dsv41_tiny and olmoe_tiny_c prints no auto-tune: block. coli doctor --json reports "tune": {} for all three.
  • glm_tiny still prints DRAFT=0 and COLI_CUDA_PIPE=1, and doctor still reports DRAFT and PIN_GB.

Not in this PR: #1584 fixes the same class of problem in coli tune's candidate sweep.

🤖 Generated with Claude Code