fix(hpa): pause HPAs via scaleUp selectPolicy instead of breaking scaleTargetRef

[arthlr]

Problem

disable_hpa() disables an HPA during migrations by renaming its scaleTargetRef to a non-existent wm--disabled--kube-* Deployment. Two defects:

• The HPA controller emits FailedGetScale / AbleToScale=False, which ArgoCD's HPA health check reads as Degraded — every migration flips the app to Degraded (observed fleet-wide during the 2026-07-08 rollout), training humans and alerting to ignore Degraded and health-gating ArgoCD wave progression/PostSync hooks.
• Under GitOps the rename is unreliable: ArgoCD owns scaleTargetRef and reverts it on any sync.

Fix

Pause with the native mechanism instead: spec.behavior.scaleUp.selectPolicy: Disabled. The HPA stays Healthy (no error events, no Degraded) but cannot scale up — the only thing a migration needs to prevent. The pre-pause spec.behavior is saved in the wiremind.io/pre-pause-scale-behavior annotation and restored exactly on resume (or the field is cleared when there was none, avoiding server-defaulted leftovers).

Safety trade-off handled: unlike the rename (accidentally healed by GitOps), a leftover pause is invisible to ArgoCD under server-side apply and would linger after a crashed run — so start_pods() and restore_stopped_pods() now resume HPAs unconditionally for every EDS-tracked Deployment (resuming an unpaused HPA is a no-op). The rename-repair path (re_enable_hpa) stays one more minor for HPAs left broken by a pre-7.0.1 run — TODO removal in 8.0.

Versioning

7.0.1 as a fix: no chart contract change (ensure_safe_run compares major.minor, chart appVersion stays 7.0.0), consumers pick it up through chartreuse~=7.0.0. No release train needed.

Validation

• 100 unit tests pass (pause idempotence, exact behavior restore, unconditional resume, rename-repair shim).
• Canary rails available if wanted: tag v7.0.1-dev.1 → PyPI dev package → staging-demo migration; expected observation: app stays Healthy during migration, pause annotation + scaleUp: Disabled visible on HPAs while the job runs, both gone after PostSync, zero FailedGetScale events.

🤖 Generated with Claude Code