feat: capability-aware scheduler + AgenticTaskWatcher + stub executor (Foreman v0.1 M2)

[Defilan]

What

Closes the Pending -> Scheduled -> Running -> Succeeded dispatch loop for the Foreman v0.1 add-on. Adds the capability-aware scheduler to the foreman-operator; the AgenticTaskWatcher, Executor abstraction, and StubExecutor to foreman-agent; and the foreman.v1 Result envelope they share. End-to-end demoed against kind-llmkube-local.

Why

Refs #500.

M2 is the architectural validation milestone for Foreman: it proves the CRDs from M0 and the FleetNode heartbeat from M1 actually compose into a working dispatch loop, without yet depending on the M3 native agent loop. The StubExecutor stands in for M3's real native agent loop so the plumbing is observable end to end today.

M2 also locked down the function-calling substrate decision the v0.1 plan rests on. Smoke-tested the live Carnice-Qwen3.6-MoE-35B-A3B-APEX-MTP endpoint serving on Apple Silicon Metal: 13/15 multi-turn flows pass cleanly; the 2 failures are a known upstream llama.cpp issue (ggml-org/llama.cpp#22072, comment with our repro added) where the strict tool-call argument parser rejects truncated JSON from the model. M3's native loop will treat this as a recoverable transient with bounded retry, getting effective success to ~99.78%. No change to M2.

How

Scheduler (internal/foreman/controller/agentictask_controller.go) evolved from the M0 logging stub into the real reconciler:

• Normalizes empty phase to Pending so subsequent logic only branches on enum values it owns.
• Cascade-fails (phase=Failed, reason=UpstreamFailed) when any dependsOn target is in phase=Failed.
• Waits with requeue while any dependsOn target is pre-terminal.
• First-fit FleetNode picker: alphabetical-by-name over Ready nodes whose capability satisfies the task's RequiredCapability (accelerator family, MinRAMGB <= AvailableRAMGB, MinContextTokens <= MaxContextTokens, NodeSelector subset of node labels). Skips nodes whose status.currentTask is non-empty (v0.1 worker-concurrency=1).
• Watches FleetNode in addition to AgenticTask; re-enqueues all Pending tasks on each FleetNode event so a node going Ready dispatches immediately rather than waiting for the requeue-after timer.

Node-side watcher (pkg/foreman/agent/watcher.go):

• Polls every --task-poll-interval (default 5s) for AgenticTasks where status.assignedNode == myNodeName && phase == Scheduled.
• Claims via status merge-patch with optimistic concurrency; race losers see the new phase on the next poll.
• Runs Executor.Execute in a goroutine; v0.1 keeps one task per node in flight via a mutex'd inflight slot.
• Re-fetches the task before the terminal patch to avoid clobbering concurrent edits. On executor error, patches phase=Failed with verdict=INCOMPLETE and the error in the Completed condition. Defensive contract-violation path on nil error + nil Result.
• Returns ErrWatcherStalled after MaxConsecutiveFailures (default 3) List() failures in a row, mirroring pkg/agent/watcher.go's supervisor-restart pattern.

Executor abstraction (pkg/foreman/agent/{executor,executor_stub,result}.go):

• Executor interface: Kind() string + Execute(ctx, task) (*Result, error). Small interface so M3 (native agent loop), M4 (gate-job executor), and any future kind plug in behind the same shape.
• Result is the foreman.v1 envelope shared between executors and any downstream consumer (planner evaluator, future DecisionLog, the human reviewer): SchemaVersion, Kind, Verdict, Summary, Extra, ElapsedSec.
• AgenticTaskVerdict constants added to api/foreman/v1alpha1 (GO, NO-GO, INCOMPLETE, GATE-PASS, GATE-FAIL, GATE-ERROR). No CRD schema change; the enum validator was already on the type.
• StubExecutor: sleeps for --stub-sleep (default 10s) and returns a synthetic GO-verdict Result. Used to validate dispatch end-to-end today; M3 swaps in the native agent loop behind the same interface.

Binary wiring (cmd/foreman-agent/main.go):

• New flags: --task-poll-interval, --task-namespace, --stub-sleep.
• Runs the registrar (M1) and the watcher (M2) concurrently via errgroup. On clean shutdown via SIGTERM both return nil; binary exits cleanly.

Tests: the M0 stub-smoke test in agentictask_controller_test.go is rewritten to exercise the M2 contract (Pending → Scheduled with capability matching, cascade-fail, dependency waits, busy-node skip). Existing M0+M1 tests untouched. Envtest harness from #501 carries forward.

Verification

Live demo on kind-llmkube-local with foreman-operator + foreman-agent (--heartbeat-interval=3s --task-poll-interval=2s --stub-sleep=8s) + apply of examples/foreman/m2-stub-task.yaml. Observed the full conditions trail in order:

Scheduled  True  FleetNodeAssigned   scheduled to FleetNode "m5-max"
Running    True  Claimed              claimed by m5-max
Completed  True  ExecutorSucceeded    stub executor slept 8.001s ...

Final status.result:

{
  "schemaVersion": "foreman.v1",
  "kind":          "stub",
  "verdict":       "GO",
  "elapsedSec":    8.001,
  "extra":         { "taskKind": "freeform", "agentName": "stub", "modelRef": "" },
  "summary":       "stub executor slept 8.001s on task default/m2-stub-demo"
}

make test passes on both arches. make lint 0 issues on darwin and on GOOS=linux (addresses the cross-arch gap captured in #503). No regressions in the LLMKube core inference flow.

Checklist

• Tests added/updated (M0 stub test rewritten as M2 scheduler test)
• make test passes locally (both arches)
• make lint passes locally (both arches)
• Commit messages follow conventional commits
• All commits are signed off (git commit -s) per DCO
• Documentation updated — README + chart README land alongside M6

What's next

M3 (#500): build the native agent loop in Go using OpenAI function calling, land the Agent CRD, ship the minimax-coder Agent. The retry policy for the upstream llama.cpp tools-strict-parse 500 (#22072) is the only architectural decision M3 takes on top of what M2 already proves works.