chore: sync claude config — hooks canon, permission guardrails

[ArthurBernard]

Summary

• Resync .claude/hooks/ with the canonical copies in ~/.claude/hooks/: drops the legacy/ exclusion from pr_decision_guard.py (the legacy tree was retired) and un-hardcodes the trading_bot label in session_start.py (read from workflow.json).
• .claude/settings.json: first-pass permission guardrails for a live-trading repo — deny Read on .env* (API keys), ask before running the real-broker pytest -m network suite. The global dontAsk mode stays; these narrow rules take precedence.
• CLAUDE.md "Common conventions" mirror resynced with ~/.claude/CLAUDE.md: model policy is now "inherit the session model"; dated mirror marker added so /groom-docs can detect drift.
• .gitignore: ignore .claude/settings.local.json + local Claude state in-repo (was only covered by a machine-global git ignore).

Why

• 2026-07-04 audit of the cross-repo Claude config. Trading_Bot is the one repo where fully-silent permissions are risky (real money, real keys), hence the narrow deny/ask rules. Note the .env deny covers the Read tool, not cat via Bash — it is a guardrail, not a boundary.

Test plan

• Config/docs only — no package code touched; python -m pytest and ruff check trading_bot/ unaffected.

[ArthurBernard]

Follow-up commit on this branch: the model policy in the CLAUDE.md mirror evolved from "inherit the session model everywhere" to tiered execution — judgement (orchestration, review, pick/plan/finish) stays on the session model; plan-leaf execution runs at the tier derived from the leaf's complexity (low→haiku / medium→sonnet / high→session model), escalating one tier on failed tests/verification.