atperson: scope/atom file convention

Dropped every underscored file name in atperson. The tree now follows <scope>/<atom>.{c,h,cpp,hpp} throughout: the directory is the scope, the file is a single-word atom.

The core already half-followed this — src/core/graph/store.c, src/core/persistence/encode.c — but the edges leaked: portable_io.c sat at the core root, scope-private headers were action_internal.h/ledger_internal.h, the CLI had graph_inspection.cpp, and all 26 test suites were something_test.c in one flat directory.

Now: src/core/io/portable.{c,h}, scope-private headers are internal.h inside their scope, tests live in tests/<scope>/<atom>.c(pp) across 20 scope directories, fuzzers in fuzz/<scope>/<atom>.c. CMake targets went kebab-case too (atperson-core-test, snapshot-load-fuzzer).

One wrinkle the rename surfaced: with internal.h in both src/core/ and src/core/action/, a same-directory include from an action atom shadows the core internal. Atoms that genuinely need core internals (candidates.c, plans.c — they read graph->node_count) now include ../internal.h explicitly, which reads as exactly what it is: reaching one scope up.

Also landed earlier in the day: the e2e lifecycle harness (issue #28) gained a conversation-context regression scenario — reply and quote context riding the durable pipeline, surviving restart from disk, quote text provably never entering the vocabulary, and a pinned boundary: context survives snapshots but not a replay rebuild, since the ledger payload is canonical text only. Same shape as the valence boundary; a ledger context section is the follow-up before #27 can rely on context through recovery.

No behaviour changes. 36/36 on both build configs.

Commit: 55c6c72