docs: proposal 77 — cache-aware model handoff architecture

[82deutschmark]

Summary

Builds on proposals #73 (complexity rubric) and #74 (UX modes) — both already merged.

Proposal #73 defined when to switch models. This proposal defines how to switch without destroying the prompt cache.

The core problem

Naive model switching (changing the model parameter mid-session after 100K+ tokens of context) costs more than staying on Opus. The cheaper model's cache starts cold and must re-process everything. The math is counter-intuitive and unforgiving.

The solution

Cache-safe subagent handoff:

1. Current model completes its work
2. Produces a structured handoff summary (compact, curated context only)
3. New subagent on target tier starts fresh from the small handoff — cheap cold start
4. Parent session cache is never touched

What's in this PR (docs-only)

docs/proposals/77-cache-aware-model-handoff.md

• Explains why mid-session model switching is wrong (with cost math)
• Defines the handoff message schema (JSON)
• Maps the complexity rubric to handoff trigger conditions
• Covers upward routing (escalation) patterns
• Covers tool set stability during handoff
• Lists anti-patterns PlanExe must avoid
• Maps to existing Luigi pipeline + MCP server architecture
• Proposes cache hit rate metrics to track

Depends on

• PR docs: task complexity scoring + model routing proposal (post-mortem + rubric) #102 (merged) — complexity scoring rubric
• Claude Code Engineering Blog: Prompt Caching at Scale

For Simon's Review

Does the handoff schema cover what the Luigi pipeline needs? Should defer_loading stubs be specified here or in a separate tools proposal?