fix(sf): add depth calibration to plan-milestone prompt

Mirror plan-slice + research-slice + research-milestone: 3-tier
Calibrate Depth (Deep / Targeted / Light) with explicit Light tier
authorizing 1-2-slice decompositions for focused well-scoped work.

Prevents the synthesized over-decomposition pattern where every
milestone produced 4-5 slices regardless of scope.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/resources/extensions/sf/prompts/plan-milestone.md

@@ -16,6 +16,16 @@ You are the first deep look at this milestone. You have full tool access — exp
 
 After you finish, each slice goes through its own plan → execute cycle. Slice planners decompose into tasks. Executors build each task. Your roadmap sets the strategic frame for all of them.
 
+## Calibrate Depth
+
+Read the milestone title, the user's stated intent, and any inlined research above. Ask: does this milestone span unfamiliar territory, multiple risky integrations, or genuinely ambiguous scope? Or is it a focused well-understood feature decomposable into a handful of standard slices?
+
+- **Deep planning** — multi-area milestone, novel architecture, multiple viable decomposition strategies. Explore broadly, write the full strategic frame, plan multiple slices with risk-driven ordering and explicit cross-slice dependencies.
+- **Targeted planning** — known territory but new to this codebase, or moderate complexity. Explore the relevant areas, plan 2–4 slices, write the requirement-mapping but skip the long architectural narration.
+- **Light planning** — well-scoped feature using established patterns. Plan **1–2 slices** with clear demos. Don't synthesize 4 slices to fill a template; use the Single-Slice Fast Path below if applicable. Skip Risks/Architecture sections that don't apply.
+
+An honest "this milestone is one slice, here's the demo" beats a fabricated 5-slice decomposition for work that doesn't have it.
+
 ### Explore First, Then Decompose
 
 Before decomposing, build your understanding: