Release v0.3.6

[youssofal]

Summary

Release v0.3.6 as the production patch over v0.3.5.

This branch carries the bounded-memory/OpenCode fixes, ports Tune into the packaged CLI, fixes verified-default onboarding/model labeling, and hardens mtplx bench tune so chip diagnostics show the exact model path and generation-window telemetry when available.

Core Pillars

• Decode TPS: protected by the existing runtime KPI and focused MTP depth tests; no release change intentionally alters draft/verify semantics after the measured memory fix.
• Prefill/TTFT: bounded KV reservation preserves prompt-context allocation and only caps huge initial new-token reserve.
• Memory: AIME-shaped max_tokens=65536 no longer reserves the full decode window up front, and anonymous one-off sessions do not retain full-capacity live cache refs.
• CLI UX: checked through actual packaged commands, including mtplx, mtplx-tune, OpenCode, Pi, and bench tune dry-run paths.

User-Facing Changes

• mtplx tune, mtplx-tune, and mtplx bench tune are available from the release package.
• mtplx start verified default now points at the installed Optimized Speed/Q4 artifact instead of prompting a bogus install.
• Tune advice is shown before measurements start, not after the benchmark has already run.
• bench tune prints exact model source notes, has --no-telemetry, waits between candidates, and scopes telemetry to generation windows when samples land inside decode.
• README now avoids claiming the speed multiplier is hardware-independent.

Validation

• python3 -m compileall -q mtplx tests scripts
• uv run --extra dev python -m ruff check
• uv run --extra dev python -m pytest -q
• uv run --extra dev python -m twine check dist/*
• scripts/fresh_venv_smoke.sh
• git diff --check
• mtplx --version -> mtplx 0.3.6 (0.3.6)
• mtplx start opencode --dry-run --json --model models/example --yes
• mtplx start pi --dry-run --json --model models/example --yes
• mtplx-tune --model models/not-loaded-in-dry-run --dry-run --yes
• mtplx bench tune --model models/not-loaded-in-dry-run --dry-run --json --yes --no-telemetry

Real Hardware Evidence Already Run On This Branch

• bench tune against /Users/youssof/.mtplx/hf-upload/Qwen3.6-27B-MTPLX-Optimized
• D3 selected at 54.51 tok/s, 2.22x AR
• D3 telemetry: scope=generation, gpu=72.0W, GPU=98.4%, window=3.5s
• Fans restored to auto afterward.

Known Non-Claims

• This PR does not claim the unrelated draft PR test: lock in real-world 3-consecutive-tool_call streaming regression #62 is ready.
• Issue Reproducing the 2.24× speedup: documented MLX patch alone doesn't get there — please publish the actual fork commit #64 remains a release-copy honesty topic; README claims were tightened here, but unavailable fork/patch proof is not invented.
• Issues GLM-4.7 PRISM loads and passes inspect after fixing num_nextn_predict_layers, but MTP generation crashes with AttributeError: LanguageModel has no attribute fa_idx #65, Add PFlash/DFlash Support #56, and Gemma compatibility #16 are not fixed by this release unless their paths are incidentally covered by existing tests.