Sessions rerouted from pinned claude-fable-5[1m] to Opus 4.8

What's happening:
My model is pinned to claude-fable-5[1m] in user-level ~/.claude/settings.json (no project-level override, no ANTHROPIC_MODEL env var). Despite the pin, my sessions are repeatedly switched to claude-opus-4-8 — with no notice, no explanation, and no indication of which layer made the switch. This has happened across multiple sessions, not as a one-off.

What I expect:
Sessions stay on the pinned model. If a fallback is unavoidable, I expect (a) a visible in-session notice that the model changed and why, and (b) a fallback within the same family — plain claude-fable-5 if the [1m] variant is capacity-constrained — not a drop to a different model tier.

Why I suspect safety routing may be misfiring:
My workload is a multi-agent orchestration / CI-CD rig: work-order pipelines, code review, release gates, registries, Slack status notifications. Nothing touches biology, security research, or offensive tooling. However, the project vocabulary is adversarial-sounding — "adversarial verify," "fault injection," "designated adversary," "kill/spawn agents," "attack the finding" — all standard software QA and process-orchestration terms. If Fable 5's dual-use safety measures are keying off this language to reroute sessions to 4.8, they're false-positiving on a plain software-engineering workload.

Questions:

1. Is the 4.8 switch capacity fallback for the [1m] variant, or safety-based rerouting? These need different fixes on my end.
2. What is the intended fallback chain for claude-fable-5[1m]?
3. Can model switches be surfaced in-session so users know when and why they're not on their pinned model?

Setup: Claude Code VS Code extension, Windows 11 Pro (10.0.26200), model pinned via user-level settings.json → "model": "claude-fable-5[1m]".