atperson: evidence-gated recall

Implemented issue #58: evidence-gated recall and tunable planning abstention.

atp_graph_recall now takes an optional atp_recall_config — a minimum overlap threshold, a most-recent-N episode prefilter bound, and a disable flag. The default config reproduces the eager behaviour byte-identically: any nonzero overlap is eligible, all episodes scanned. An optional atp_recall_report returns the evidence for each call: episodes total/scanned/matched/returned and the gate that applied (none, disabled, prefilter, min-overlap), the same carry-the-evidence contract as action candidates.

The gate is policy, not learned state — no vocabulary, familiarity or network mutation, no snapshot bump. The prefilter scans the most recent episodes because episodes are stored in insertion order; it bounds the scan, not the ranking within it.

The guarded decision layer already read its thresholds from atp_action_decision_config; the gap was CLI surfacing. decide now takes them positionally — min-candidate, min-support, max-drop, max-consecutive — and prints the effective thresholds with every decision. plans prints its planner config. recall takes gate args and prints a recall-report line.

Also fixed CI: the Linux GCC build had failed since the worker-pool commit because parallel.cpp used std::condition_variable without including <condition_variable> — libc++ pulls it in transitively, libstdc++ doesn't.

• gate tests: default byte-identity, disable, gate excluding all matches, partial matches, prefilter bound, repeated deterministic calls
• 46/46 tests on both build configs