Spike: DSM-backed shared memory — #21 cannot free RAM under static pg_shmem_init!

[styk-tv]

All of pgRDF's shared state is statically allocated at postmaster start:
pg_shmem_init! over PgLwLock / PgAtomic in src/storage/shmem_cache.rs,
with the dictionary cache a fixed [DictCacheSlot; SLOTS]. Two consequences:

1. pgrdf must be in shared_preload_libraries or the shared surface is inert
   (src/lib.rs:50 gates on
   process_shared_preload_libraries_in_progress; src/storage/stats.rs:39
   notes the counters read zero without it).
2. The allocation cannot shrink. It exists for the life of the postmaster.

Why this blocks #21 specifically

#21 (C4 · park / unpark graph — seal lifecycle) promises "free + reload
RAM/CPU/disk." Under the current model, parking a graph cannot return the
dictionary-cache RAM — the slots are sized once, at start, and stay. #21's
headline promise is not reachable as specced without a dynamic allocation
mechanism.

pgrx has carried storage/dsm_*.h bindings since 0.17.0. Dynamic shared memory
would let the dict cache and the staged-loader job control allocate and release
per-graph, which both makes #21's "free" real and removes the
shared_preload_libraries requirement as a hard precondition.

Scope — this is a spike, not an implementation

• Establish whether DSM / dsm_registry can back the dict cache and
  staged/jobctl without regressing the hit rate or the concurrency model.
• Establish what park / unpark would actually reclaim, measured.
• Decide whether C4 · park / unpark graph — seal lifecycle (free + reload RAM/CPU/disk) #21 keeps its current shape or is re-specced.

Output is a written finding on this issue, not a merged design.

Why P2

#21 is P4. The spike only needs to precede #21's implementation, not this run.

Relates

Blocks the "free" half of #21. styk-tv/pgCK#20 §3 records the same
shared_preload_libraries coupling on that side, where it is satisfied today but
becomes load-bearing as the identity and presence work lands.