Finalize telemetry at invocation completion - #14390

Builds on #14360; review against that branch to isolate this alternative.

Description

Keep one attachment event, without calling back into command state during telemetry shutdown.

#14360 combines the number of supplied attachments with the Markdown operations completed by their uploads. Those facts become available at different times. Its RecordDeferred callback builds the final event during Flush.

This alternative records the facts as they become known. The telemetry service copies them into a pending event; Main still owns completion, using one deferred Finish() to snapshot and send the payload. Attachment commands use a small, typed interface:

opts.AttachEvent = attachments.BeginTelemetry(
	telemetry, cmd.CommandPath(), opts.AttachFlag.Count(),
)
opts.Assets, err = opts.AttachFlag.UserAssets()
// Handle validation errors.

// After UploadAndAttach, before handling any upload error:
opts.AttachEvent.RecordOperations(uploadResult)

Intentional difference from #14360: attachment telemetry begins immediately before file validation, not in Args. Supplying 51 attachments still records attach_count=51 when the 50-file limit rejects them. Authentication, positional-argument, and other failures before UserAssets() do not produce an attachment event or promote attachment sampling. Generic command_invocation telemetry now covers those earlier failures.

How did you test this change?

I built and recorded gh at 38704130e876 with GH_TELEMETRY=log, isolated configuration, and dummy credentials. The recordings show real commands; the visible sed/jq pipeline selects fields from their actual telemetry logs.

The upload/write scenarios use a local, non-forwarding simulated GitHub API, reached through isolated http_unix_socket configuration. They exercise the real client and Markdown transformation, not GitHub's production upload or comment service.

Rejected input still counts

I supplied 51 attachment values to gh issue comment. It exited with the 50-attachment limit error, while recording attach_count=51, zero operations, and sample_rate=100. This case needs no API calls.

Two references are one replacement

I supplied two PNGs, referenced the first twice in the body, and left the second unreferenced. I saw both references rewritten and the second image appended in the actual outgoing comment request. Telemetry recorded two attachments, one replacement, and one append at full sampling.

I also injected failures into the same local fixture:

Both failure cases retained the paired command event and full sampling. Separately, during development I exercised all six commands with positional errors and earlier flag conflicts, plus an unauthenticated issue-comment invocation: each logged only the generic command event, not an attachment event.

Authorship and follow-up

Who wrote this:

• A human wrote it.
• An agent wrote it under close human direction.
• An agent wrote it independently, and no human has guided the implementation beyond the initial prompt.

Who answers review comments:

• @williammartin will read and reply directly.
• An agent will draft replies and @username will read them before they are posted.
• Nobody has explicitly committed to replying.