paths: follow-ups from #542 (LRU order/NULL pinning tests, per-chunk lruMu) #548

Follow-ups from the #542 review (Macmini, APPROVE with nits), none blocking:

1. N1. The LRU eviction order changed from handler candidate order to sorted observation id. Same entries and same hit/miss decisions, but when one request inserts past lruMaxSize (10000) a different subset survives. Pin the intended order with a test, or document it.
2. N2. Nothing pins that an absent row or a NULL resolved_path must not enter the LRU; dropping the present guard survives the suite.
3. N3. Removing both sort.Ints calls survives every test, so the chunk determinism they provide is unverified.
4. N4. Make the longestTx mapping structurally safe with a seen set or map[int][]int. This is defensive only; it is not reachable through ingest today.
5. N5. lruMu is taken once per id inside a chunk, up to 500 round-trips; take it once per chunk instead.

Relates to #542.