feat(mecak8s): enable in-chat schedules for remote execution

[JAORMX]

CI status / integration blocker

Head: adb733ff1806f37fae8d0c8edd970daac949889a.

• Vercel deployment completed successfully; its check watcher finished.
• GitHub reports CONFLICTING / DIRTY against current main. A non-mutating merge-tree check identified internal/app/scheduler_fire.go, overlapping the scheduler cancellation/ownership lifecycle changes merged in fix(scheduler): keep polling behind stuck fires without releasing ownership #2277 (8e3bea42e).
• No GitHub Actions run was created for this head: pull-request Actions are blocked by the merge conflict. This is not green CI and does not replace the local validation limitations below.
• No rebase, force-push, branch merge, or conflict resolution was performed. Updating the branch requires operator approval, followed by integration validation and a fresh CI watch. The PR remains draft.

Summary

Enable in-chat scheduled tasks that borrow a native Kubernetes session's exact remote workspace. Restore the remote Schedule/ScheduleQuery catalog and remove the implicit CLI scheduler disable; explicit --no-scheduler remains honored.

• Reauthorize the exact owner, origin binding, environment ref/revision and placement scope on each fire, then reserve and publish a distinct successor reference for that fire session.
• Complete run cleanup with FinishRun, defer attachment reservation closure until publication, release process-local resources after ambiguous Commit responses without deleting durable state, and best-effort abort the exact reservation after a lost ReserveSuccessor response.
• Add factory, scheduler, authenticated controller-backed and real Kind regression coverage. Update the canonical scheduled-task documentation and native-execution/architecture pages.
• Fix the Kind fixture's proven API-egress failure: development previously omitted runtime discovery even though kindnet enforces NetworkPolicy. Discover and validate exact service/endpoint IPv4 /32s and API ports for both profiles; keep workload fixture egress production-only. Before/after TCP probes and a passing original-image baseline established the cause; no authorization, fencing or workload isolation was weakened.

Supported scope and intentional limits

• Workspace-backed remote schedules must be created in chat and borrow the origin workspace. Independent remote workspace allocation is unsupported and fails before allocation; independent no-FS schedules remain supported.
• A busy workspace fails the fire under the existing trigger policy; this does not introduce durable waiting or broaden retries.
• The origin session must remain stored with the same binding. Schedule updates cannot retarget it. Retained fire sessions consume the existing shared 64-reference limit; deleting a schedule alone does not free those references.
• Reservation abort remains best effort, not a distributed orphan-free cleanup guarantee. Ambiguous Commit retains the durable snapshot/reference and does not run the model.

Validation

This is deliberately a draft: the single full local race gate did not complete within 600 seconds, and live qualification is limited to the borrowed-schedule scope below.

Validation	Result
Final-source task --concurrency 4 lint	PASS
Full task --concurrency 4 test, before the final narrow reservation-response repair	PASS; 191 passing package results
Full task --concurrency 4 test:race	TIMEOUT at 600 seconds, no reported failure/race before cancellation; not claimed as a pass
Combined bounded race runs before the narrow repair	Complete 218/218 unique package outcomes, plus three tagged passes; not a single successful full race invocation
Final-source executionclient, executioncontroller, server and app packages with -race -count=1	PASS
Narrow repair regression	Failed before the repair, passed afterward, including race detection
task docs and task site:build after the narrow repair	PASS
Offline go run ./cmd/mecademo, before the narrow repair	PASS; tool result and permission ask/approval demonstrated
Final-source TestKindExecutionBorrowedSchedule	PASS, no skips, 341.03 seconds

The final live test required two completed manual fires and a future durable timer after both provider and agent restarts. Every fire required successful Shell/Read event replay and completed persisted snapshots, distinct published fire references, exact environment/PVC continuity and remote-file verification. It used a scripted mock LLM, not paid model calls. This is not a claim that the entire production-network/renewal suite ran.

An additional explicitly tagged e2e lint invocation reported five findings outside the final narrow repair (context-argument ordering and shadowing, including shadowing in the new e2e helper). Ordinary full lint passed; that extra tagged invocation is not claimed green.

Reproducible committed coverage

Offline regressions:

go test -count=1 ./deploy/mecatl-execution-kind ./cmd/mecak8s
go test -race -count=1 ./internal/adapter/executionclient ./internal/adapter/executioncontroller ./internal/adapter/server ./internal/app

The committed live test is TestKindExecutionBorrowedSchedule in e2e/k8s_execution/borrowed_schedule_test.go (build tag kind_execution_e2e). The fixture command below builds/loads the images and tagged test binary, establishes explicit owned cluster context and fixture identity, and runs the production-profile qualification including this test:

CONTAINER_ENGINE=podman \
MECATL_EXECUTION_DEV_TOOLBOX=dev \
MECATL_EXECUTION_K8S_TOOLBOX=sre \
MECATL_EXECUTION_QUAL_CI=0 \
MECATL_EXECUTION_QUAL_NATURAL_EXPIRY=0 \
task e2e:k8s:execution:production

This is a reproduction entry point, not a claim that this entire command passed during the final focused validation. It provisions a new fixture and retains it locally; do not use it to mutate the retained validation cluster. A focused rerun with the fixture's explicit state/context and authenticated helper environment uses -test.run '^TestKindExecutionBorrowedSchedule$' -test.v -test.count=1 -test.timeout=10m on its tagged binary.

The existing native executor deployment is retained, with all 20 environment/Pod/PVC identities preserved and no deletion or unfencing. Final deployed agent: localhost/mecak8s@sha256:908dad40c1c3cac8a19c722b0ab003417080f8306ea5b8d0d03a67c6fe4a5aa3. Local receipts, fixture state, artifacts and credentials are not committed.

API compatibility

Backward-compatible, additive availability of the existing scheduling API in remote-execution deployments. No protobuf/HTTP schema, wire major, feature identifier, engine public Go API, or persisted schema changes. Existing conforming clients need no migration; deployment-specific availability is not protocol version skew. The explicit scheduler opt-out remains supported. Internal service plumbing changes do not alter the engine's public Go surface. Stored records remain compatible and exact placement authorization stays fail-closed; persisted tool grants are not silently widened. Older sessions without Schedule authority require a new eligible session to expose it. No breaking change or major-bump approval is required.