fix(agents): deny apply_patch for GPT models to prevent verification hangs (#2935)

GPT models (5.3-codex, 5.4, etc.) frequently hang when using apply_patch
due to verification loops. This adds:

1. Tool restriction: apply_patch is denied for GPT variants of
   Hephaestus, Sisyphus-Junior, and Sisyphus agents
2. Prompt guidance: GPT-specific prompts now explicitly instruct using
   edit/write tools instead of apply_patch
3. Removed the 'Always use apply_patch' instruction from
   sisyphus-junior/gpt-5-4.ts that contradicted the fix

The deny is model-conditional — Claude variants retain apply_patch
access since it works reliably there.

src/agents/sisyphus/gpt-5-4.ts

@@ -304,7 +304,7 @@ Every implementation task follows this cycle. No exceptions.
    Skills: if ANY available skill's domain overlaps with the task, load it NOW via \`skill\` tool and include it in \`load_skills\`. When the connection is even remotely plausible, load the skill - the cost of loading an irrelevant skill is near zero, the cost of missing a relevant one is high.
 
 4. EXECUTE_OR_SUPERVISE -
-   If self: surgical changes, match existing patterns, minimal diff. Never suppress type errors. Never commit unless asked. Bugfix rule: fix minimally, never refactor while fixing.
+   If self: surgical changes, match existing patterns, minimal diff. Never suppress type errors. Never commit unless asked. Bugfix rule: fix minimally, never refactor while fixing. Use the \`edit\` and \`write\` tools for file changes. Do not use \`apply_patch\` on GPT models - it is unreliable here and can hang during verification.
    If delegated: exhaustive 6-section prompt per \`<delegation>\` protocol. Session continuity for follow-ups.
 
 5. VERIFY -