fix(transcribe): preserve saved output on write failure

[rudycelekli]

Problem

Both transcript output paths truncate the destination before writing: the CLI uses Path.write_text, and the standalone podcast script redirects its final report directly to the output. A mid-write failure returns nonzero but destroys a previously complete transcript.

A native OS file-size ceiling reproduces this in both paths, including an output symlink: the old complete output becomes partial data. Upfront path failures already return nonzero; this fix addresses loss of existing output after writing starts.

Changes

• Assemble/write the complete transcript before replacing the destination, using a temporary file beside the resolved target.
• Follow existing output symlinks to their target while preserving the link and existing regular-file permissions.
• Keep previous data intact when staging, flushing or replacement fails, clean temporary files, and report a nonzero failure.
• Require regular-file output targets so devices/FIFOs/directories are preserved rather than replaced.

The managed shell script remains standalone; it uses only the existing Python dependency. API, media, interpreter and polish behavior are unchanged. Atomic replacement creates a new file inode, so hard-link/owner metadata is not preserved; this does not promise recovery from every crash/power-loss scenario or directory fsync durability.

Validation

• Before:4 failing native file-limit cases,4 successful-save controls; both CLI and shell paths and symlink targets covered.
• After:11 focused cases pass, including portable UTF8 file replacement, permission preservation on POSIX, failed replacement cleanup, FIFO preservation/nonblocking refusal and staging cleanup.
• Full suite:618passed.
• Ruff all passed; bash syntax/diff checks clean.
• mypy: same3 pre-existing MCP SDK2.2 diagnostics.
• Format: same54 existing files, no additional failures.

No real podcast/Groq requests. The exact failure is a process file-size ceiling; a full-disk failure is an inference from the same write-error class, not a separate reproduction.

AI assistance: prepared with Codex and reviewed before submission.