/model selection silently overridden (Fable 5.1 → Opus 4.8), likely a security-classification false-positive on a benign Wine game-compat task

## Summary

An explicit model selection made via `/model` was silently overridden mid-session. I selected **Fable 5.1** (and saved it as my default), but the session continued to run as **Opus 4.8**. I did not choose Opus 4.8 at any point.

## What I observed (verifiable)

- I ran `/model` and set the model to **Fable 5.1** (twice — the second time it was saved as my default for new sessions).
- Despite that, the session's own status line reported **"You are powered by the model named Opus 4.8."**
- So my explicit `/model` choice was not honored; the effective model was something I never selected.

## What I expected

`/model` is an explicit user choice. If the harness must route away from the selected model (for policy/classification reasons or otherwise), it should at minimum be **visible and explained** — not a silent override. Ideally the user's selection should be honored, or the user should be told *why* it can't be, with a way to opt back in.

## Suspected cause (my hypothesis — not confirmed from inside the session)

The session was a long, benign task: getting a **legally-owned game client (MapleStory TW) to run on Linux under Wine**. The work involved things a naive security classifier could easily flag:

- Binary-patching Wine's *own* DLLs (`kernelbase.dll`, `win32u.so`) to fix Wine compatibility bugs
- Reading/analyzing anti-cheat components (BlackCipher / NEXON Game Security) from logs
- Disassembly and socket/`WINEDEBUG` tracing
- Vocabulary like "bypass" — used **only** to state that bypassing anti-cheat was explicitly *refused*

None of this is offensive security. The task never touched credential theft, malware, exploitation of third-party systems, or anti-cheat circumvention (that was explicitly declined). If a security classifier is what triggered the model change, this is a **false positive** on legitimate software-compatibility / reverse-engineering-for-interoperability work.

I want to be precise: I can confirm the *override* (selected Fable 5.1, ran as Opus 4.8). I cannot confirm from inside the session that a security classification *caused* it — that is my inference based on the task content.

## Why this matters

Reverse-engineering and binary patching for **interoperability / getting owned software to run** is a mainstream, legitimate developer activity. Silently downgrading/overriding the user's chosen model on these tasks (a) breaks the explicit `/model` contract, and (b) penalizes benign work. At minimum the override should be transparent; ideally the classifier should not treat "make my game run on Linux via Wine" as a security threat.

## Repro

1. Start a session and use `/model` to select a model (e.g. Fable 5.1).
2. Do a task involving Wine DLL binary-patching + anti-cheat log analysis + disassembly.
3. Observe the effective model reported by the session become/remain a different model (Opus 4.8) than the one selected.

## Environment

- Claude Code, Linux (Ubuntu, kernel 6.8), interactive terminal session
- Model explicitly selected: Fable 5.1; effective model observed: Opus 4.8

Start by reproducing the /model selection flow in Claude Code on Ubuntu and compare the selected model with the session status line during the described task. No source file or test is named; done means the selected model remains effective, or any routing override is visible and explained to the user.

Written by the indexing model from the issue text.