Modded multiplayer finish-line investigation and candidate correction

HISTORICAL BUILD RECORD
The original findings and limits below belong to the named earlier build.
For current instructions, use docs/README.md and CHANGELOG.md in the
source repository. These old results do not qualify later packages.

===================================================================

Observed failure
----------------
The supplied host/client reports run the same Experimental Rollback engine,
US v1.0 ROM, two-owner roster and admitted mod namespace. Both enter the custom
race on retail carrier 5 successfully. At the first finish the host stops at
epoch 92/frame 2320; the client continues to 2338 before the session ends.
The host reports that its bounded rollback receive queue is no longer serviced.
Neither supplied report contains a native crash stack; both return to the
launcher. The last "finished" counter changes from 1 to 2, but that diagnostic
must not be treated as an authoritative count of completed racers.

Correction
----------
The confirmation owner previously called SDL audio-device functions and durable
EEPROM/controller-Pak storage synchronously. Device locks, frequency changes,
write-through flushes and backup rotation can therefore stop packet consumption.
Those blocking calls are now separated from the packet-consuming thread:

* Confirmed audio is copied to an eight-block, fixed-size mailbox. The ordinary
  event owner consumes it after input polling. No guest spans survive a callback.
  If a device stalls, only old already-confirmed PCM is discarded, with sparse
  diagnostics; gameplay inputs, effects/checkpoints and peer hashes are retained.
* Confirmed saves are copied to a dedicated writer with one active write and one
  latest pending snapshot. Existing validation, isolated online paths, leases,
  atomic replacement and backup/history rules remain in force. The final
  confirmed snapshot must finish before the match-return handshake releases the
  route. Write failures remain errors, never a fallback to an offline save.
* Native owner-stage diagnostics identify a slow capture, restore, tick, effect
  confirmation, publication, scene load or durable write. Exceptions racing a
  connection end are now logged instead of losing their underlying diagnostic.

The supplied logs do NOT identify the exact blocked call, so this is a candidate
correction to demonstrated blocking paths, not a claim of an end-to-end verified
finish-line fix. No race logic, mod definitions, queue admission limits, rollback
prediction rules or legacy backends are changed by this correction.

Qualification
-------------
Headless native tests exercise copied buffer ownership, invalid audio blocks,
audio ordering/overflow, concurrent producer/consumer operation, blocked save
writes, bounded latest-snapshot coalescing, final durability fences, write
failure and exception reporting. These checks do not replace a real modded
multiplayer finish-line test. No game windows or scripted gameplay are launched.
Windows and Linux builds completed successfully. All 16 selected native
regression suites passed on each platform; the new confirmed-output suite also
passed 100 consecutive repetitions on each platform. Packaged executables are
verified against the rebuilt binaries, with no ROMs, saves or crash logs shipped.

Install this candidate on BOTH host and clients. It retains the Playtest 14
version name; use the packages in the finish-output-fix-20261006 directory rather
than an older Playtest 14 download. Test an AI first finish and a player first
finish on the previously failing custom track. If the session still ends, retain
both runtime.log files, including any [rollback][owner-stall] and
[rollback][retired-operation] lines.
