docs(claude): codify model-selection + PR-labeling rules surfaced in cross-repo work

[cristim]

Summary

Three rules that were absent from the §2 Subagent Strategy section, each surfaced by a real failure mode during a batched-PR-shipping session on a different repo:

1. Explicitly set the model parameter on EVERY Agent call

Omitting the parameter causes the subagent to inherit the parent session's model tier. When the parent is Opus, every subagent silently becomes Opus — a 5–10× cost premium for work that almost never warrants it. The fix is simple: always pass model: explicitly, even when you want the parent's tier (making the choice visible at the call site).

2. Routine PR-shipping defaults to Sonnet end-to-end

The standard plan + 3-pass review + worktree + implement + test + push + open-PR + ping-CR + arm-CI-watcher pattern is Sonnet-tier work. When acceptance criteria are clearly enumerated, Sonnet's review catches the right things and its implementation passes verification cleanly. Escalate to Opus only when a specific design judgment is needed — not as a default for "the task feels important."

3. Every gh pr create MUST be followed by mirroring triage labels onto the PR

PRs that lack triage labels are invisible to the same priority queries that surface issues. Treating label-mirroring as a post-open-PR follow-up (rather than part of the open-PR step itself) means it gets dropped, especially in subagent delegation where the prompt doesn't mention it. The rule specifies which labels to mirror, how to handle PRs that close multiple issues, and that subagent prompts must include this step explicitly.

Acceptance criteria

• All three bullets are present in the §2 Subagent Strategy section of CLAUDE.md
• Placement is before the existing model-tier rubric (Match model to task complexity…)
• No existing content is modified or removed