🧠 refactor(claude.md): verify-context doctrine + tighter trigger + MCP merge

[ZaxShen]

Summary

Three changes per user line-by-line audit on PR #129:

1. Trigger rule compressed (14 lines → 8)
2. MCP + agent model merged (12 lines → 4) — full agent model already in docs/AGENTS.md
3. NEW: 'Before answering — verify context' section (10 lines) — two concrete checks bro runs before any substantive answer

The verify-context doctrine

User wanted explicit anti-guessing rules with two concrete questions:

• 'Is this the industry standard or the best way?'
• 'Do you have enough context from DB, codebase, user's ideas?'

Web-researched (see machinelearningmastery.com hallucination tricks, promptfoo prevent-llm-hallucinations) and polished into:

## Before answering — verify context

**Don't guess. Don't fabricate. Don't be a yes-man.** Before you plan, decide, or answer a substantive question, run two checks:

1. **Context check** — *do I have enough?* Pull from this priority order: codebase (Read / Glob / Grep / git), DB (MCP queries), web (WebFetch / WebSearch for upstream docs / specs / standards), then training-data fallback. If context is thin, **say so** and either ask the Human or run the lookup. Thin context → 'I'm not sure, checking…' beats inventing an answer.
2. **Standards check** — *is what I'm about to recommend the industry standard or the best way?* If you're not sure, do the lookup. If a domain expert (legal, security, perf, etc.) would handle it better than bro, propose `tmb_agent-creator` to spawn the specialist. Bro should be professional and competent across general SWE work; for genuinely specialized domains, escalate.

When you're guessing, label it. Cite the source when relevant.

Replaces the implicit 'don't be a yes-man' that lived only in tmb_concerns-protocol (which only fires on disagreement). This new section fires on every substantive answer.

Subagent CLAUDE.md isolation (defense in depth)

Per user concern about permission leakage, added one line to agents/swe.md and agents/pr-reviewer.md:

Do not read project-level CLAUDE.md — that file is bro's persona; this agent's prompt is canonical for [SWE / review] work.

CC does NOT auto-load CLAUDE.md into subagent context (subagents get their own system prompt from agents/<name>.md). This line is belt-and-suspenders against a subagent inadvertently Reading the file from disk and following bro-only rules.

Followups (filed separately)

• Q1: A/B prompt testing in L5/L6 — to be filed as new GH issue
• Q6: per-project AGENTS.md auto-update + name-collision strategy with user's local agents

Local verification

• agent-line-budget passes (swe 23, pr-reviewer 28; cap 30)
• CLAUDE.md: 112 → 103 lines (cumulative origin → here: 268 → 103, 62% reduction)