fix: remove preemptive context warnings in plan-eng-review

[claygeo]

Problem

The plan-eng-review skill's priority hierarchy instruction tells the model to watch for low context:

"If you are running low on context..."

This causes premature compression warnings on 1M context window models. At 27% usage on a 1M window, ~730K tokens remain — far from any real constraint. The model doesn't have precise token-counting ability and relies on heuristics, so this instruction makes it "context-anxious" regardless of actual available space.

Context Used	200K remaining	1M remaining	Should warn?
27%	~146K	~130K	200K: maybe. 1M: absolutely not
50%	~100K	~500K	200K: yes. 1M: no
80%	~40K	~200K	Both: yes

Solution

• Replace the self-monitoring trigger with system-level and user-initiated triggers only
• The model now compresses when: (1) the user explicitly asks, or (2) Claude Code's own compaction system fires
• Explicitly instructs the model not to preemptively warn about context limits

Changes

• plan-eng-review/SKILL.md.tmpl — Updated priority hierarchy instruction
• plan-eng-review/SKILL.md — Regenerated from template

Test Plan

• Run /plan-eng-review on a 1M context model with a long conversation — verify no premature compression warnings
• Run on a 200K context model — verify compaction still works when system triggers it
• Verify user can still say "compress" to trigger manual compression

Closes #510