Add names-only format to search_beliefs for 6.5x token savings

Start in tools.py at _search_beliefs() and inspect the BELIEF_TOOLS schema for the existing search_beliefs definition. Update the documented system prompts in EDITOR_SYSTEM_PREFIX and PROMPT_SYSTEM_PREFIX, then verify that full remains the default and names-only returns only the truth value and belief ID.

Summary

Add a format parameter to search_beliefs supporting names-only mode. Belief names carry 90%+ of functional value at 11% of token cost (E_coord evidence). For a 64K context window, this is a significant budget savings.

Current behavior

search_beliefs returns [IN] belief-id: first 120 chars of text -- approximately 130 tokens per result. With 20 results, that's ~2,600 tokens (4% of a 64K context window).

Proposed change

Add format parameter to search_beliefs:

Format	Output	Tokens per 20 results
full (current default)	[IN] belief-id: text[:120]	~2,600
names-only	[IN] belief-id	~400

Names-only uses 6.5x fewer tokens for the same search. The bee can search more broadly and then use show_belief for the specific beliefs it needs full text on.

Why this matters for worker-bee specifically

Worker-bee runs small models (Qwen 27B) on small context windows (64K). Every token matters. Names-only search lets the bee:

• See 6x more results per search (orientation)
• Spend context budget on investigation rather than belief text
• Use show_belief selectively for the 2-3 beliefs it actually needs

Where names-only works well for small models

• The bee's own brain beliefs -- the bee chose the names, so they're maximally informative to it
• Well-known programming concepts -- atomic-write-requires-fsync, redis-uses-pubsub. A 27B code model has these in its parameters
• Orientation searches -- "what does this EEM know about X?" Names give the landscape cheaply

Where names-only needs follow-up for small models

• Codebase-specific jargon -- keeper-sprinkling-prevents-tool-pair-split needs a show_belief call for a 27B model that hasn't seen the source
• A 27B model has less parametric knowledge than Opus/Sonnet, so more names will need description follow-up. But even with 50% needing show_belief, total tokens are still lower than loading all descriptions upfront

System prompt update

Add to EDITOR_SYSTEM_PREFIX / PROMPT_SYSTEM_PREFIX:

When searching beliefs, use format="names-only" for broad orientation.
Use show_belief for full text on specific beliefs you need to reason about.

Evidence

• E_coord: names-only Haiku +55pp, full beliefs all-at-once -5.8pp
• 0.615 median cosine similarity between name and description embeddings (same neighborhood)
• 900x compression ratio for names-only vs source documents
• See: beliefs-pi ftl-reasons#265 for the corresponding reasons CLI changes

Implementation

Minimal change to _search_beliefs() in tools.py:

def _search_beliefs(query, limit=20, format="full"):
    ...
    for r in all_results:
        if format == "names-only":
            lines.append(f"[{r.get('truth_value', '?')}] {r['id']}")
        else:
            lines.append(f"[{r.get('truth_value', '?')}] {r['id']}: {r.get('text', '')[:120]}")

And add format to the tool schema in BELIEF_TOOLS.