Model list is hardcoded; allow custom/newer model IDs, and "Opus 4.6 (1M)" now fails roadmap generation #2034

Title: Model list is hardcoded; allow custom/newer model IDs, and "Opus 4.6 (1M)" now fails roadmap generation

Version / OS: Auto Claude 2.7.6, Windows 11, bundled Python 3.12.8, Claude Code 2.1.318 (Max plan, OAuth profile)

1. Roadmap generation fails with opus-1m

• Roadmap model set to opus-1m (Claude Opus 4.6 (1M)).
• UI shows "Roadmap generation failed with exit code 1"; the app log has no roadmap error.
• Reproduced by running runners/roadmap_runner.py --model opus-1m --thinking-level high with the app's env. The runner prints
  There's an issue with the selected model (opus-1m). It may not exist or you may not have access to it. and then
  Agent did not create discovery file x3, and exits 1.
• --model opus and --model sonnet run normally on the same account.
• opus-1m maps to claude-opus-4-6 plus beta header context-1m-2025-08-07 (phase_config.py MODEL_BETAS_MAP). I did not confirm that the header is the cause.
• Ask: surface the runner's real error in the UI/log, not just the exit code.

2. Newer models cannot be selected

• The dropdown (AVAILABLE_MODELS), MODEL_ID_MAP (main bundle) and phase_config.py MODEL_ID_MAP only know Opus 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5.
• I patched in claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5, claude-fable-5-1. All four answer claude -p --model <id> on the same profile.
• Ask: let users enter a custom model ID per feature/phase (the backend already passes unknown IDs through), or load the list from config/claude instead of hardcoding it.

3. Adaptive thinking / effort is tied to one model ID

• ADAPTIVE_THINKING_MODELS is {"claude-opus-4-6"} (backend) and ["opus","opus-1m"] (UI). Newer models are silently treated as non-adaptive.
• Ask: make this a capability flag per model, not a hardcoded set.

4. Updates overwrite local changes

• autoUpdateAutoBuild replaces resources/backend on update, so any local model-map change is lost. A user override file for models would avoid this.