feat(cli): let transcribe --to write srt or vtt straight from a media file - #5215

What changed

hyperframes transcribe video.mp4 --to srt used to stop with "--to can only export from transcript files. Run transcribe first.", so getting captions from a recording took two commands. Now --to srt or --to vtt on an audio or video file transcribes it and writes the sidecar in the same run, next to transcript.json (or at --output).

• Transcripts keep working as before: --to on a .json, .srt or .vtt still exports instead of importing, with the same cue rules.
• Both paths share one sidecar writer, so cue grouping, --preserve-cues and --output behave the same.
• Under --json the run still prints one result line; it gains format and outputPath when a sidecar was written.
• A bad --output fails at once, before any decoding, so a typo does not cost a long transcription. That covers a missing folder ("Output folder not found") and an --output that is a folder or ends in a slash ("--output is a folder", naming a file path to use instead). A caption path that would land on the input file or on the transcript.json the run writes is refused too ("The caption file would overwrite ..."). The check compares files, not path text: right before the caption write, by when the transcript exists, a destination that is the same file as the input or transcript.json is refused, so the operating system resolves every link, relative link, folder alias (/var and /private/var) and letter case. An early check refuses the common cases before transcribing. When a disk reports inode 0 (some network shares), existing files are compared by their fully resolved paths instead, so links are still caught and two different files are never called the same. It also covers the default path, so transcribe transcript.srt --to srt no longer rewrites its own input. A trailing -o with no path fails ("--output needs a file path").
• If the caption file cannot be written after transcribing, the run exits 1 and says so: "Transcript saved to ..., but the caption file ... could not be written: ", instead of reporting a failed transcription.
• Transcribed words are treated as word-level when grouping cues, rather than guessed from spacing. For Chinese or Japanese this gives readable caption lines instead of one cue per recognised word. Exporting the saved transcript.json later still guesses from spacing, so it may split those cues differently; the docs say so.
• Audio with no speech succeeds with 0 words as before, and with --to writes an empty sidecar (an empty SRT, a header-only VTT). Exporting an empty transcript file still fails with "No words found".
• With --optional, a run skipped because whisper-cpp is unavailable writes no sidecar and exits 0. Before, --to on media always failed with exit 1.
• The CLI docs and --help examples show the new form.

Validation

• New tests: a media file with --to vtt --json writes transcript.json, a VTT that matches the formatter's output, and exactly one stdout result naming both; a missing --output folder fails before transcription starts; silent audio with --to srt writes an empty sidecar. The first two fail on the previous code. New tests also cover an --output that is a folder, ends in a slash or sits under a file, an --output that names the input or transcript.json directly, through a link, through a linked folder, by its real path or in other letter case, an export whose default path is its own input, a trailing -o, a caption file that cannot be written after transcribing, and CJK grouping. Each new test fails when its fix is removed. Links to the transcript the run is about to create (direct, and relative inside a linked folder) keep the transcript and report that the caption file could not be written. All 63 transcribe tests pass three runs in a row. On macOS, the built-from-source CLI refuses each alias (case variant, /private path, link, trailing -o, export onto its own input) and leaves the files untouched.
• End to end with the built CLI and the real Parakeet engine on a 3.5 s generated speech clip: transcribe hello.wav --to srt --json exits 0, prints one result line, and writes three correct cues ("Write HTML.", "Render video.", "Built for agents.").
• transcript.json is named by one constant, used by both engines, the import, init and the check.
• Format, lint, type check and comment checks pass.

No visible change

CLI only; no Studio or player code changes.