feat: add cache-aligned prompt builder for LLM API cost savings

PROBLEM:
- PromptBuilder.build() puts everything in a single user message
- System + query + memory + chunk all change together
- LLM prompt caching gets 0% hits (entire message differs per call)
- For a 50-chunk ingestion run, we pay full input price 50 times

SOLUTION: PromptBuilder.build_cache_aligned()
- Splits prompt into 3 separate messages:
  1. SYSTEM: instructions (stable across ALL calls) → CACHED
  2. USER[0]: query/problem (stable per run) → CACHED
  3. USER[1]: memory + chunk (varies per call) → not cached
- Cache prefix (system + query) reused across all chunks in a run
- Estimated 30-70% cache hit ratio depending on chunk sizes
- ~50% input token cost savings for multi-chunk ingestion

TEMPLATES:
- templates/gru-mem-system.txt (instructions only, 840B)
- templates/gru-mem-query.txt (problem wrapper, 29B)
- templates/gru-mem-turn.txt (memory + section, 57B)
- templates/gru-mem.txt (legacy, unchanged)

API:
- PromptBuilder::build() — legacy, backward compatible
- PromptBuilder::build_cache_aligned() → PromptMessages
- PromptMessages.cache_prefix_tokens() — cacheable token count
- PromptMessages.total_tokens() — total estimated tokens
- PromptMessages.headroom() — tokens available for response

TESTS: 11 unit + 3 integration = 14 new tests
- test_cache_aligned_produces_two_user_messages
- test_cache_prefix_is_stable_across_chunks
- test_cache_prefix_is_stable_across_memory_changes
- test_cache_prefix_tokens_positive
- test_headroom_positive_under_budget
- test_legacy_build_still_works
- test_cache_aligned_contains_query
- test_cache_aligned_memory/chunk_budget_exceeded
- a8_cache_prefix_stable_across_50_chunks
- a9_cache_aligned_headroom
- a10_cache_savings_estimate

TOTAL: 64 mem-core tests passing (52 unit + 12 integration)

templates/gru-mem-query.txt

@@ -0,0 +1,3 @@
+<problem>
+{prompt}
+</problem>

templates/gru-mem-system.txt

@@ -0,0 +1,10 @@
+You are a memory gate for a recurrent update system. You evaluate whether new evidence sections contain information relevant to a standing problem, and update the running memory accordingly.
+
+Rules:
+1. Retain all relevant details from the previous memory while adding new useful information.
+2. Judge whether you have collected enough information to answer the problem.
+3. Reason about the new section between <think> and </think>.
+4. If the section contains useful information: output <check>yes</check>, then update memory between <update> and </update>.
+5. If the section does NOT contain useful information: output <check>no</check>, then keep previous memory unchanged between <update> and </update>.
+6. If more information is needed: return <next>continue</next>.
+7. ONLY when enough information is collected: return <next>end</next>.

templates/gru-mem-turn.txt

@@ -0,0 +1,7 @@
+<memory>
+{memory}
+</memory>
+
+<section>
+{chunk}
+</section>