Remove unused legacy prompt helpers (init_prompt.go), delete empty
branch and unused slugRegexp in name_step, remove unused formatDomains
in egress_step, and run gofmt across all TUI source files.

UI/cli ux update

- Add web_search provider abstraction with Tavily (default) and Perplexity backends
- Auto-detect provider from env vars (TAVILY_API_KEY > PERPLEXITY_API_KEY),
  with explicit override via WEB_SEARCH_PROVIDER
- Add tavily-search as external skill with embedded script and registry entry
- Add `forge skills add` command for vendoring registry skills into projects
- Update init wizard with provider selection phase when web_search is chosen
- Send Telegram "typing" chat action while LLM processes requests

- Add CLAUDE.md with pre-commit gofmt and golangci-lint instructions
- Wire validateTavilyKey into the init wizard tools step
- Fix gofmt formatting in skills.go and tools_step.go
- Rename ValidatePerplexityFunc to ValidateWebSearchKeyFunc for both providers

…alidation

- Egress step now only shows the selected web_search provider domain
  (Tavily or Perplexity) instead of both
- Skills step prompts for required_env, one_of_env, and optional_env
  when skills are selected, blocking if required/one_of is not provided
- Web search tool step validates Tavily/Perplexity API keys with async
  validation (same pattern as provider step)
- DeriveEgressFunc now receives envVars to filter provider-specific domains
- Updated tests to match new provider-specific behavior

Tools line now displays "web_search [tavily]" or "web_search [perplexity]"
to indicate which provider was selected.

The review step reads from WizardContext directly, not the tools step
Summary(). Updated Prepare() to annotate web_search with the provider
from ctx.EnvVars.

Init() was not resetting phase, currentPrompt, envPrompts, or envValues.
When the user navigated back from a later step, the stale skillsEnvPhase
with an exhausted currentPrompt caused an index out of range panic.

Skills step now checks env vars collected by earlier wizard steps
(via Prepare/knownEnvVars) so it won't re-prompt for keys like
OPENAI_API_KEY already provided as the model provider. buildEnvVars()
uses a written map to skip duplicate keys when writing the .env file.

Provider step Apply() now stores the API key in ctx.EnvVars
(e.g. OPENAI_API_KEY) so the skills step's Prepare() can detect
it via knownEnvVars and skip the one_of group entirely.

MultiSelect and SingleSelect Init() methods were no-ops that didn't
reset the done flag. After navigating back, done remained true causing
Update() to short-circuit and treat all keys (including space) as
no-ops while the step immediately re-completed.

Three issues fixed:
- Wizard now handles esc globally as quit (matching kbd hints)
- Removed redundant esc→back mapping from review step that caused
  the review↔egress loop
- EgressDisplay.Init() now resets done flag so back-navigation
  doesn't immediately re-advance

Add Tavily web search integration and Telegram typing indicator