feat(sweep): run isolation fans concurrently on the device-farm fleet

[jonathaneoliver]

Summary

Run a sweep isolation fan (control + single-axis-flip variants) concurrently across the Appium Device Farm — one device per arm, all sharing the isolation group/pattern — instead of serially on one operator-pinned sim. This makes cross-platform/protocol flips (the design's Tier-1 isolation strategy) actually parallel and cuts wall-clock per fan.

Motivation

sweep's IsolationFan (tools/harness-cli/internal/sweep/isolate.go) already builds a control + one variant per axis-flip (OFAT attribution). But TestSweepProbe (modes/sweep_probe_test.go) resolves one operator-pinned device (PickMode → appium.Discover → CHARACTERIZATION_DEVICE_UDID) and never touches DeviceFarmEnabled()/resolveFleetDeviceFarm() — so the fan runs serially on one sim. The skill explicitly wants the opposite: "Tier 1 = platform/protocol → different devices → simultaneous."

char-matrix already has the machinery: TestCharMatrixFleet (modes/char_matrix_fleet_test.go) → resolveFleet → resolveFleetDeviceFarm (modes/fleet.go) allocates N devices by capability, assigns per-arm platform (CHAR_ARM_<i>_PLATFORM), pins real-iPhone hardware off-farm (runner/appium.go:287), and binds all arms to one proxy group (binding is by player_id at the proxy). Post-#869 both probes converge on the typed charplan.RunPlan (go-proxy/pkg/charplan/plan.go).

Proposed

When a fan has ≥2 arms, route it onto the fleet path: build a charplan.RunPlan from the fan and resolve via resolveFleetDeviceFarm, so control + variants launch on one device each, sharing the isolation group/pattern (control arms the pattern, variants bind). Keep the single-device path as the default for depth-1 / single-experiment ticks (cheap, one sim — the loop's cost-per-tick discipline).

Scope

• Teach the sweep probe a fleet entry point — prefer reusing the TestCharMatrixFleet path via charplan.RunPlan over duplicating allocation logic.
• Fan → RunPlan: each arm = one ArmConfig (recipe + PlayerID + shared Group/PatternMaster).
• Gate: fans (≥2 arms) → fleet; singletons → existing single-device probe.

Key files

internal/sweep/isolate.go (IsolationFan), cmd/harness/sweep.go (apply/agenda), modes/{sweep_probe_test.go,fleet.go,char_matrix_fleet_test.go} (probe + DF allocation), go-proxy/pkg/charplan/plan.go (RunPlan), runner/appium.go (hybrid real-iOS routing, withBaselineTestFlags).

Constraints / gotchas

• Mixed-platform fans (sim + real iPhone) use the appium.go mixed-master design: pin the iPhone hardware UDID, leave sims unpinned (fleet.go:117). The iPhone arm runs off-farm on :4799, farm-locked while sims stay on the farm at :4723 — the ADF cannot WDA-launch a real iOS 26 device (see memory reference_real_iphone_usbmux_pairing_wipe). The fan must tolerate one arm on the hybrid path.
• Preserve Fleet char-matrix runs have no graceful-interrupt teardown → orphaned appium sessions block the next run #853 device-release-on-interrupt (t.Cleanup/SIGTERM dance) and the n=1 confirm-reps gate.
• Keep the queue fileless / CH-master — thread the RunPlan in-memory / temp-file (as char.go already does); don't introduce a spec file just to reach the fleet path. (If the YAML-bridge sibling lands first, the fleet path may reuse its RunPlan adapter.)
• Stale device-farm locks after an abrupt kill must still be reaped (the documented reapDeviceFarm / unblock recovery).

Verification

• Unit: fan → RunPlan golden — arm count, shared group, single-axis diffs preserved.
• Live: seed a platform-flip notable → control + variant launch on two devices at once, share one pattern (identical bandwidth per tick), both land verdicts → analyze attributes correctly. Plus a mixed sim + real-iPhone fan exercising the off-farm hybrid arm.
• Build gate: go build ./... && go vet ./... across go-proxy, tools/harness-cli, tests/characterization.

Acceptance criteria

• A ≥2-arm isolation fan launches control + variants on separate farm devices concurrently, sharing one group/pattern.
• Cross-platform and mixed sim+real-iPhone fans both run (iPhone off-farm, farm-locked).
• Depth-1 / single-experiment ticks still take the single-device path.
• Fleet char-matrix runs have no graceful-interrupt teardown → orphaned appium sessions block the next run #853 release-on-interrupt + confirm-reps gate intact.

Relationship

Sibling of the sweep↔YAML-bridge issue, which introduces the Experiment ↔ charmatrix/RunPlan adapter this path reuses. Land the YAML bridge first; this builds on its RunPlan mapping.