refactor(agent): derive recall guidance from the tool set, drop the flag

[flarexium]

Problem

The cross-turn recall feature had two switches that had to agree:

• the prompt's Recall rules block, gated by PromptRenderer.RecallEnabled;
• the recall tools (recent_entries / get_entry), wired off the agent's RecentEntries buffer.

They were flipped together at each call site (buildBookEngine(..., recall) + the agent's Recent buffer), but nothing enforced agreement — a future change could leave the prompt advertising a recall ability the tool set doesn't back, and the model would call tools that aren't there.

Change

Make the tool set the single source of truth. PromptRenderer.Render derives the recall guidance from whether input.Tools advertises recent_entries — which the harness already populates every turn from the runner's tool registry (loop.go sets cycleInput.Tools). Wiring the recall tools is now the only thing that turns the guidance on.

Removed:

• PromptRenderer.RecallEnabled field
• the recall bool parameter on buildBookEngine (and the true/false at the three call sites)

No behavior change

• TUI wires the recall tools (agent carries a RecentEntries buffer) → guidance on.
• book-run / bench carry no buffer → no recall tools → guidance off.

TestPromptRenderer_RecallRulesDerivedFromTools now drives the prompt off input.Tools instead of the flag. go build ./..., go vet ./..., go test ./... all pass.

🤖 Generated with Claude Code

[jarvis-flarex]

LGTM