fix(plan): respect explicit COLI_CUDA_MTP=1 - skip the planner's DRAFT=0 export (opt-in was silently inert)

[mohamedmastouri2000-boop]

What

Per the request in #467: when COLI_CUDA_MTP=1 is explicitly in the environment, _auto_tune skips its DRAFT=0 export so the engine's DRAFT=-1 auto path (colibri.c — g_draft = (has_mtp && (!cuda || cuda_mtp)) ? 3 : 0) can actually see the opt-in. Plus the ENVIRONMENT.md line next to COLI_CUDA_MTP.

Why

The documented opt-in was silently inert on the Windows bare-run and --auto-tier flows: resource_plan.py _auto_tune exported DRAFT=0 (compute-bound, or disk-bound with projected hit < 0.90), which preempts the engine's auto path — the only place COLI_CUDA_MTP is consulted. Anyone A/B-ing speculation under CUDA got a silent no-op unless they also knew to set DRAFT=3 (my first A/B pass in #467 measured exactly baseline-vs-baseline noise because of this).

Behaviour

• COLI_CUDA_MTP unset (default): unchanged — planner still exports DRAFT=0 on compute-bound / low-hit-disk plans; MTP stays off under CUDA. This is the measured-correct default ([Performance]: 5.42 tok/s — dual-socket Xeon Gold 6430 AVX-512 VNNI, 1TB DDR5 RAM #389 compute-bound −42%; [Performance]: RTX 5080 / 128 GB resource matrix (drive x compute x RAM) ΓÇö RAM dominates; COLI_CUDA=0 Windows quirk; PR-ready N-drive shard split #467 streaming-bound −32%).
• COLI_CUDA_MTP=1 explicit: planner leaves DRAFT alone → engine auto path resolves draft=3. An explicit user DRAFT=n continues to win over everything (environment_for_plan is setdefault-based).

Verification

On the #467 box: COLI_CUDA_MTP=1 alone (no DRAFT) through the bare-run Windows flow now loads with MTP ACTIVE (draft=3) and speculates (2.00 tok/forward on a 2-token smoke); without the var, draft=0 as before. The doc line also records the measured trade-off so nobody reads the opt-in as free speed: acceptance holds (49% on sm_120) but at ~85% hit the widened S=4 expert union costs −32% net; the opt-in pays near full residency.

Scope

c/resource_plan.py (one guard + comment in _auto_tune), docs/ENVIRONMENT.md (one row). No engine changes.

[JustVugg]

Merged into dev — thank you @mohamedmastouri2000-boop. Verified locally beyond CI: unit-tested _auto_tune both ways — COLI_CUDA_MTP=1 at disk/0.85 hit now leaves DRAFT unset (opt-in engages draft=3), and unset still exports DRAFT=0 (the measured-correct default). import os present, CI 8/8.

This closes the third-and-last wrapper gotcha from #467 — the documented opt-in now actually works, so anyone can reproduce your MTP-under-CUDA A/B (49% acceptance, −32% at 85% hit) without the DRAFT=3 incantation. One tiny non-blocking note for a future pass: COLI_CUDA_MTP=1 set on a no-GPU box would now also skip the DRAFT=0 export (it's a CUDA flag, so that's user-error territory) — if you ever want to tighten it, gating on has_gpu would scope it to where it's meant to apply. Not worth a PR on its own; just noting it.

With this, #467 is fully closed: findings 1/2/4 shipped (#468/#469), finding 3 measured with the default confirmed, and the opt-in now honoured (#470). Genuinely one of the best-run issues in the repo.