trace2: improve lifecycle and instrumentation - #2408

Consolidate event serialization and normal/performance formatting so each
message declares only its event-specific data. Extract the shared format and
process classifications, and align writer implementations with their source
names to make later lifecycle changes smaller and clearer.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

TRACE2 represents one process-wide event stream, so an injected
instance obscures its lifetime and makes region state appear isolatable
when it is not. Make event emission and writer state static first so
the API shape can be reviewed apart from startup ownership and the
downstream constructor migration.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Tracing begins before CommandContext exists and survives until
dispatcher shutdown, so injected services create a circular dependency
for process identity and settings. Let Trace2 establish its SID, depth,
targets, and initial events at entry, then close writers after dispatch
ends.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

CommandContext should expose replaceable execution services, not own a
process-wide event stream. Removing TRACE2 from that surface makes the
remaining injected dependencies explicit and avoids preserving a fake
test tracer for state that can no longer vary by context.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Core infrastructure cannot meaningfully substitute a process-wide tracer,
yet carrying it through process, Git, HTTP, and platform constructors
forces every helper to mirror global lifetime. Remove that dependency
from the service layer and emit its existing events through the static
API, while leaving leaf exception conversion for the final pass.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

OAuth clients only forward the tracer so every provider can emit into
the same process stream. That parameter creates a constructor cascade
without representing per-client state. Remove it at the OAuth boundary
so provider wiring follows the real ownership model, while reserving
leaf exception changes for the final cleanup.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

After constructor ownership is gone, leaf exception wrappers and command
call sites still refer to a tracer that cannot vary by caller. Route
those final authentication, provider, storage, UI, and Win32 events
through the process API so the migration becomes buildable again
without introducing an adapter that would immediately disappear.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Exception subclasses that emit while being constructed can report failures
more than once and make ordinary error handling depend on tracing. Emit once at
the application boundary, preserve the richer Git and interop formats there,
and return call sites to standard exception types.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Child start and exit events need a stable correlation identity and timing
that survives fast process termination. Give each child wrapper its own Trace2
identifier, classify it when it is created, and report the operating system
exit timestamp from the process notification.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Physical managed threads do not describe asynchronous work once execution
crosses awaits or dispatcher boundaries. Flow an explicit Trace2 context with
the execution context so thread identities, nested regions, and relative timing
remain coherent, while allowing callers to restore the main context when
needed.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Repeated fully qualified caller-info attribute names obscure the
Trace2 API signatures. Import the namespace to keep those signatures
readable without changing behavior.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

A process start records argv but does not identify which command parser
branch ultimately won. Emit the canonical command name after parsing and carry
an inherited hierarchy through child processes so nested helper activity can
be related to its originating command.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Regions describe elapsed work but cannot explain the values that shaped
that work. Add scalar and structured data events whose thread, nesting, and
relative timing come from the active logical context, allowing later
instrumentation to attach useful measurements without inventing new regions.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Top-level traces currently stop before command parsing and provider
selection, leaving the most important dispatch decisions invisible. Add focused
regions and metadata around setup, input parsing, provider resolution, and each
command entry point so command execution can be followed end to end.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Command dispatch alone cannot explain time spent discovering Git, reading
configuration, or opening the selected credential store. Trace those service
boundaries and record the configuration scope and backing-store choices that
drive their behavior.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Eager console creation and dispatcher hand-offs hide UI startup cost and
can initialize terminal state before it is needed. Create consoles lazily and
trace ANSI, Avalonia, rendering, and window-display boundaries so interactive
latency is attributed to the correct lifecycle stage.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>

Authentication spans browser, device-code, refresh, account selection,
and provider-specific decisions that are otherwise indistinguishable in a
trace. Add regions and mode data at those boundaries so delays and user-flow
choices can be diagnosed without recording credentials.

Assisted-by: GPT-5.6 Sol
Signed-off-by: Matthew John Cheetham <mjcheetham@outlook.com>