Owned typed skinning results (experimental mode3)

native_skinning_result.h separates the portable buffer contract from the FMA
implementation. native_skinning_result.cpp accepts this contract, a calculation
version and a pure callback. Inputs are complete original positions, influences
and the declared palette span. Outputs are transformed positions and per-vertex
frames. Guest addresses, GPRs and local-store scratch are not identity or output.

On a miss, an exclusively leased slot owns copies of all inputs and a separate
output area. The native callback computes directly into that output storage.
The result remains available after the invocation. On a hit, every input byte,
size, arithmetic control and version is checked before copying outputs to the
current invocation's destinations. A content hash only selects candidates.
Actual guest publication, completions and final-pair state reconstruction remain
in the title adapter. This is a bridge to persistent geometry ownership, not
yet a GPU buffer lease or a native render snapshot implementation.

Lifecycle and synchronization
32 shards,64 slots each;2 MiB buffer reservation per shard. Total buffer budget
is64 MiB, including in-flight reservations. Entry metadata, cache base objects
and allocator overhead are additional. Allocations are rounded to4 KiB; slots
reuse their allocations when capacity suffices. LRU removes resident results
and releases capacities as needed. Busy slots cannot be read, overwritten or
evicted. No lock is held while copying inputs or calculating a miss. Cache hits
hold the shard lock during exact input comparison/output copies. If no slot or
budget is available, calculate directly with no caching. Process-lifetime cache
avoids teardown races; explicit reset requires no live calls. Allocation failure
releases the reservation and falls back without touching another worker's lease.

Caller obligations
Validate every input/output span, palette index, count and alias relationship.
Keep inputs stable during the call. Outputs must not alias read-only inputs or
each other; positions are intentionally in/out. All exceptions must be masked.
Callback must return normally: no exception/longjmp/I/O, retaining borrowed
pointers or changing arithmetic control. Results must be independent of incoming
sticky FP flags. These conditions are true for the extracted skinning callback,
but are not an automatic purity guarantee for arbitrary code.

Floating environment
Identity includes all MXCSR control bits, but not incoming sticky flags. A miss
clears only the sticky flags before the pure callback, records the flags raised,
then restores incoming flags OR raised flags. A hit applies the same OR to this
call's incoming environment. The native skinning callback does not consult flags
or change controls. Thus rounding, exception masks and output bits are preserved
while equivalent geometry can be shared across callers with different flags.
Unexpected control changes decline publication of a reusable result.

Bleach mode3 preserves mode2 bounded arithmetic, the span/format/alias gates,
fresh pointer updates, and the final original pair. Exact checks compare every
GPR/local-store byte and MXCSR against the original lifted loop. The default
remains off; mode2 is the current measured experimental reference until mode3
has clean net-performance evidence. Neither counts of hits nor raw vertices
reused alone establish an FPS improvement.

Validation199
The title-derived differential corpus passed2,304 cases, including real kernel
misses/hits with changed incoming flags, extended/denormal/boundary floats and
full architectural state comparisons. The generic storage test additionally
covers all64 incoming flag combinations, mutations at the end of each input,
control/version changes, eight concurrent workers, budget pressure/eviction,
and quiescent reset. See live audit and clean results separately.

Live audit199 and net comparison200/201
Audit199 passed72,495 exact GPR/local-store/MXCSR comparisons and completed
10,337 submissions with zero pending. Six stills show character/scene/tutorial
and one fade frame; these do not approve continuous shadows or interpolation.
Same executable SHA256
14ffb227aab1dfb0901050e0be3bf7360d189b6b3518dfc57c6d48dc2892b875:
  mode2 reference201:78.6111 FPS, p9513.6884ms, p9915.8782ms;
  mode3 experiment200:79.7222 FPS, p9513.5427ms, p9916.1560ms.
Observed net difference +1.4134%,0.1773ms/frame saved; one clean run per mode,
same18-second route, not deterministic replay or statistical confidence.
Mode3 reused93.87% of batches and95.97% of its vertices, but that fraction is
not a percentage of frame cost removed. Input comparison, copies and locks
remain; the original producer and downstream geometry still execute.
Keep mode2 as the experimental reference. Mode3 is opt-in, not promoted from
this small unreplicated difference with a worse p99. Do not spend the next
optimization cycle tuning its size/hash parameters. The next contract must
span full geometry computation, persistent data lifetime through all consumers
and publication, and draw preparation; isolated skinning cannot close the gap.
Reports: outputs/Revisao-estrutural/skinning-persistente-199-201.txt/json.
