v0.15.5

Fixed

Revive the learning feedback loop. Three breaks were keeping confidence scores meaningless — 99.7% of facts had success_count=0 because the outcome pipeline was structurally dead:

• Synthesize unified_diff from code_block when the LM returns an empty diff for a real file path. Empty diffs collapsed every outcome to UNVERIFIED, and success_count only grows on ACCEPTED / linked-UNVERIFIED.
• Tighten the code-generation prompt: diff and code_block are now REQUIRED for real file paths; empty is only permitted when file_path="/". Closes the escape hatch that made ~34% of suggestions signal-less.
• Run pruning on every FactStore.initialize(), not just when outcomes fire (prune_stale_facts / demote_unhelpful / purge_dead). Inactive projects were accumulating review facts indefinitely. Also shortened the staleness window for INDEPENDENT-tagged facts to 7 days since they're 0.2-confidence noise by construction.