MUST-gap fill, batches 2-3 — JVM, DSA, distributed, LLD, and the one missing AI card
Continued filling the coverage-audit MUST gaps: 31 more /concepts cards (JVM internals, MVCC, system-design estimation/scaling, DSA, distributed-systems depth, LLD patterns) and the one genuinely-missing AI card (vector indexing). 52 cards added total across the three batches.
Kept filling the gaps the audit surfaced, worst-covered domains first. Batch 2 added 31 full bilingual cards to /concepts (156 → 187): JVM internals (memory areas, the memory model / happens-before, concurrent collections, PreparedStatement, the lock-types comparison), MVCC, the system-design method gaps (back-of-envelope estimation, vertical-vs-horizontal scaling, stateless tier, cache write patterns), the DSA holes (topological sort, finite-state-machine modeling, integer-overflow/money precision, amortized analysis, bucket top-K), distributed-systems depth (consistency models, failover/split-brain, delivery semantics, session guarantees, write-ahead log, replicas-vs-sharding, cascading failures, consistent hashing), and the low-level-design patterns (state, transition table, factory, design-for-testability, composition-over-inheritance, make-illegal-states-unrepresentable, adapter, and when NOT to apply a pattern). Each card is the same full shape as the rest of the deck and was generated then fact-verified by a second agent (75 corrections).
Batch 3 was a single card: vector indexing (HNSW / IVF / pgvector) for /ai. The audit flagged five AI gaps, but checking the existing /ai ids showed embeddings, output safety, observability, and idempotency were already covered by dedicated cards (embedding, data-exfiltration, pii-secret-handling, observer, cost-budget, idempotent-tool, retry-policy). Only ANN vector indexing was genuinely absent — so adding the other four would just have duplicated, which is exactly the lossy-duplicate problem the /concepts dedupe earlier removed. Added the one real gap with correct pgvector SQL and the recall-vs-latency-vs-memory trade-off.
Net across the three batches: 52 cards, /concepts 136 → 187 and /ai 66 → 67. The worst-covered domains (Backend/Java + concurrency, DSA, low-level design, distributed-systems depth) are now filled at the MUST level. The full map and what's still open are tracked in docs/coverage-gaps.md. Still open and worth doing: a couple of behavioral cards (trust-battery and proactive no-surprises communication — which are the interviewer's stated hot buttons), a first-principles-reasoning-aloud speaking phrase, and a few low-priority meta-skill/edge cards.