atperson: graph growth bounds

atperson e048c3e — graph growth bounds and scale benchmarks (issue #9).

The semantic graph grew without bound and every lookup scanned. atp_find_node walked the node array with strcmp; atp_find_edge walked the edge array. Both are now open-addressing hash indexes over the canonical arrays: FNV-1a token hash for nodes, a packed (source, target) key for edges, power-of-two capacity, linear probing, load factor held at or below 0.5. The arrays stay the source of truth — indexes are derived state, maintained incrementally on intern and edge creation, rebuilt by the snapshot loaders, never persisted.

The first edge-key hash was identity on the packed key. That was the bug of the day: intern indices grow together, so packed keys are dense and sequential, and identity hashing packs linear-probe runs into contiguous spans. A 76k-edge snapshot loaded in 1.8 seconds. Profiling pointed straight at the probe loop in the index rebuild. Running the key through a splitmix64-style finalizer spread it: load went to 16ms, and observe cost at 50k observations fell from 177µs to 44µs per observation — the same clustering was quietly degrading atp_find_edge during observe. Per-observation cost now scales with the pair count, not graph size.

Resource ceilings: atp_graph_config gains node_capacity_max and edge_capacity_max (0 = unlimited). An observation that would cross a ceiling is rejected whole with ATP_ERR_CAPACITY — a dry-run pass counts the new nodes and edges a text would create before any mutation, so nothing is half-learned. Rejections are counted in capacity_rejections. Ceilings are deployment policy, not graph data: atp_graph_load restores unlimited ceilings regardless of the saving process's budget, and atp_graph_set_capacity applies one to a live graph. Lowering a ceiling below the current count evicts nothing — existing state stays, further growth rejects. Pruning with provenance stays future work; episodes reference nodes by index, so deletion is a separate decision.

Benchmarks (tests/bench.c, ctest -L bench): deterministic synthetic histories at small/medium/large — 1k/10k/50k observations over 500/5k/25k vocabularies, skewed so a few tokens carry hub pressure. Observe throughput, association lookup, recall, snapshot save/load, memory estimate. Timing is reported, never asserted, so CI variance can't flake. Growth tests cover whole-observation rejection, hub nodes, one-off vocabulary bursts, index-rebuild equivalence after snapshot roundtrip, and replay equivalence under an admitting ceiling.