Remote launches were always uncapped — no wall-clock cap ever reached the cloud runner

[aarontrowbridge]

Problem — The GPU-plane architecture pass (2026-08-18) found that launch.ts constructs RemoteExecutor() bare: max_wallclock was never sent on ANY remote launch since the feature landed — no CLI flag, no solvespec field, no default. The cloud bundle has no cooperative stop and the client's lostAfterMs only marks OUR view terminal, so a wedged remote run bills until the instance dies. The exact 'hung GPU job' failure mode.

Approach — The cap is never optional: resolution ladder explicit > $AMICODE_REMOTE_MAX_WALLCLOCK_S > generous default (7200s). Payload always carries max_wallclock.

Acceptance Criteria

• Every remote submit payload carries max_wallclock (default 7200; env override respected; explicit verbatim) — test-pinned

Source

GPU-plane architecture pass 2026-08-18 (vault note pending); companion finding: the cloud runner exposes no observability surface (SKU/quota unknown pre-launch) — cloud-side ask, tracked separately.

[aarontrowbridge]

Fixed by #420 (merged).