[feat] Allow per-task model tiering instead of forcing opus on all agents

[Hai4320]

What problem does this solve?

Problem

The skill forces every agent to model: "opus" (SKILL.md:93,437,
agent-design-patterns.md:213, orchestrator-template.md). Mechanical agents
(QA static-analysis, deploy log parsing) don't benefit from Opus — it just costs
more and runs slower. And because the rule is duplicated + in the audit
checklist, any project override to sonnet gets reverted on the next sync.

Proposed solution

Tier by task, keep aliases:

• reasoning / design / code → opus
• mechanical / IO / deploy → sonnet

Soften the "always opus" wording in the 3 spots and make the checklist verify
"model fits the task tier" instead of "every agent is opus".

Alternatives considered

No response

Related team-architecture pattern (if any)

Not pattern-specific

Additional context (optional)

I realized a few problems when working on a personal project in Vietnamese.
Hai4320/hai_portfolio#32