fix: cache-friendly feed injection ordering in cllama

[mostlydev]

Summary

• Reverses feed/time injection order in cllama so the static system prompt is the prefix (cacheable) and dynamic content (feeds, timestamp) is appended after
• Fixes Anthropic prompt cache thrashing that was costing ~$60/week on the tiverton-house pod ($91.56/week → estimated ~$30/week)

Root Cause

InjectAnthropic and InjectOpenAI prepended dynamic content before the static system prompt. Anthropic prompt caching is prefix-matched (5-min TTL), so the volatile prefix invalidated the cache on every request — causing cache_creation_input_tokens at 1.25x cost instead of cache_read_input_tokens at 0.1x cost for the ~120K static system prompt.

Changed Files (cllama submodule)

• internal/feeds/inject.go — swap prepend → append in 3 code paths (OpenAI string, Anthropic string, Anthropic content-blocks)
• internal/feeds/inject_test.go — update assertions to expect system-prompt-first order
• internal/proxy/handler_test.go — update 2 handler tests from HasPrefix to Contains + ordering assertion

Test plan

• All cllama unit tests pass (go test ./... — 11 packages)
• Deploy to tiverton-house and verify cache_read ratio improves over 24h via cllama session-history JSONL analysis

Fixes #122

🤖 Generated with Claude Code