fix: stabilize Telegram GPT-5.2 reasoning replay, add facilitator skill, enable web search via openrouter/perplexity, and restrict certain functions

[Vardominator]

Summary

This PR stabilizes Telegram agent sessions on openai/gpt-5.2 by preventing context replay mutation that can orphan OpenAI Responses reasoning items, and adds repo guidance/config updates that were part of the same rollout.

Problem

Telegram conversations were intermittently failing with:

400 Item 'rs_...' of type 'reasoning' was provided without its required following item.

This was observed when replaying prior turns that included reasoning + tool-call chains.

Changes

• Disable context pruning by default in the shared OpenClaw base template.
  • templates/openclaw/org.base.json5
• Mirror that setting in the tracked local override example.
  • instances/core-human/config/instance.overrides.example.json5
• Document prevention + one-time recovery steps for already-poisoned sessions.
  • docs/CONFIGURATION_DETAILS.md
• Pass BRAVE_API_KEY into compose environment and cover it in tests/docs.
  • src/compose.ts
  • tests/compose.test.ts
  • .env.example
  • README.md
• Add a new repo skill for rollout/doc discipline.
  • skills/thoughtful-oco-facilitator/SKILL.md
  • skills/thoughtful-oco-facilitator/agents/openai.yaml

Validation

• bun test (24 pass, 0 fail)
• Reproduced tool-call + follow-up session flow on openai/gpt-5.2 after config update with no replay error
• Gateway health check passes for core-human

Operational Notes

• The config fix prevents new sessions from being poisoned.
• Existing broken sessions still require a one-time session reset/new session mapping (documented in docs/CONFIGURATION_DETAILS.md).