[Bug] Opus 4.6 max effort mode generates verbose output without meaningful execution or instruction adherence

Bug Description
Title: Opus 4.6 + max effort: 250k tokens burned with zero meaningful output on well-documented task
Description: Brand new session. Opus 4.6, max thinking enabled. Task: integrate Cosmograph engine into an existing graph component. Project has custom skills, hooks, documented API endpoints, and a UX analysis document — all explicitly referenced in the prompt.
What happened: Over ~30 minutes and 250k tokens, Claude Code:

Read irrelevant docs (Vercel) instead of researching the library I explicitly told it to research
Made a single trivial config change (link opacity 3% → 7%)
Declared 6 tasks complete with no visible changes
Took a screenshot of a blank WebGL canvas and claimed success
Generated extensive ceremony (task lists, skill invocations, summaries) with no substance

Expected behavior: Follow explicit instructions to research Cosmograph, apply findings, build interactive features documented in the UX spec.
Suspicion: Max effort appears to be causing the model to spend tokens on planning/narration loops rather than execution. The model itself confirmed it "defaulted to the lowest-effort interpretation of every instruction" despite max effort being set.
Environment: Opus 4.6, max effort, full context available, custom skills architecture.

Environment Info

• Platform: darwin
• Terminal: WarpTerminal
• Version: 2.1.86
• Feedback ID: 4e1cc0af-5415-47ed-95a6-79e4b56eb956