AI Runner Files

Runtime fixes (verified contributor round 2026-09-06)

Two verified runtime fixes, merged as byte-exact standard merge commits (no scope creep):

#2081 GuiDaemonClient.cancel_runtime raised TypeError on every call

cancel_runtime never forwarded the keyword-only metadata and timeout_seconds arguments that _runtime_action requires (its siblings load_runtime/unload_runtime did). Every call raised TypeError, and the TTS interrupt handlers except RuntimeError did not catch it — so interrupting TTS propagated the error instead of cancelling the request. cancel_runtime now forwards both arguments; auto_start keeps its existing False default so cancelling never starts the runtime. Regression test added (test_gui_daemon_client_runtime_actions.py, parametrized over all three wrappers) — 4/4 pass; the file fails on the pre-fix commit.

#2082 Quantization auto-select returned INT4 where it cannot fit

_auto_select_quantization ended with an else branch byte-identical to the elif above it, so the fallback returned the 4-bit config exactly when VRAM was below the INT4 budget (vram < 0.5× model). The else now selects QuantizationLevel.INT2 (0.25×, requires_calibration=True) — the level that was declared, configured, and supported end-to-end but never selectable through the automatic path. Verified over a 120-cell no-hardware grid: only the sub-0.5× cells changed, every explicit preferred_level= choice is untouched.

Notes

• Merges: 62763220a (#2081), 90f0c509c (#2082); version bump d969635d4.
• PR [#2072] (third-party-notice CI job) closed as superseded — its change was already on master via 01a9e3e22 (#2074); its branch carried an unrelated revert.
• Fixes commit as 983c59481 (INT2 fallback) and 96e05b5b0 (cancel_runtime forwarding).