fix(cli): improve TTS and model download reliability

[miguel-heygen]

Summary

• Add retry with exponential backoff (3 retries, 1s/2s/4s) to the shared download utility, with socket timeouts (30s), response timeouts (60s), redirect loop protection, and specific HTTP 403/429 rate-limit error messages
• Add file-level locking (.lock files) to TTS model and voice downloads to prevent concurrent processes from racing on the same temp file
• Increase hasPythonPackage timeout from 10s to 30s to accommodate ONNX runtime cold-start
• Set maxBuffer to 10 MB on the synthesis subprocess to prevent truncation from verbose ONNX warnings

Problem

The TTS command fails frequently at scale due to four compounding issues:

1. GitHub rate limiting — the 311 MB model downloads from GitHub Releases have no retry logic. HTTP 403 responses are terminal.
2. Download races — concurrent agent-spawned TTS calls write to the same .tmp file, causing corruption.
3. False negative package checks — ONNX runtime cold import exceeds the 10s timeout on constrained machines, making hasPythonPackage("kokoro_onnx") report false even when it's installed.
4. maxBuffer overflow — ONNX prints verbose warnings that exceed Node's 1 MB default, killing the subprocess.

Reproduction

# 1. Rate-limit failure (download.ts has no retry):
# Clear the model cache to force a fresh download:
rm -rf ~/.cache/hyperframes/tts/models/

# Run tts — if GitHub rate-limits the 311 MB download, it fails immediately
# with no retry. In agent environments with 3+ concurrent tts calls,
# this is near-guaranteed at scale.
hyperframes tts "Hello world" -o /tmp/test.wav
# Error: Download failed: HTTP 403

# With fix: retries 3 times with 1s/2s/4s backoff, and prints:
# "Download failed: HTTP 403 (rate limited). GitHub throttles
#  unauthenticated release downloads. Retry in a moment."

# 2. Concurrent download race (no file locking):
# Run two tts commands simultaneously — both see model as missing,
# both start downloading to the same .tmp file:
rm -rf ~/.cache/hyperframes/tts/models/
hyperframes tts "Hello" -o /tmp/a.wav &
hyperframes tts "World" -o /tmp/b.wav &
wait
# One or both fail with corrupted model file

# With fix: second process sees .lock file, waits for first to finish

# 3. ONNX cold-start false negative (10s timeout too short):
# On a cold machine or constrained sandbox:
pip install kokoro-onnx  # installed but slow to import
hyperframes tts "test" -o /tmp/test.wav
# Error: The kokoro-onnx package is not installed
# (false negative — import timed out at 10s)

# With fix: timeout increased to 30s

Scope

The download utility (packages/cli/src/utils/download.ts) is shared by TTS, Whisper, and background-removal commands — all three benefit from the retry and timeout improvements.

Test plan

• Build passes (bun run build)
• All pre-commit hooks pass (lint, format, typecheck, fallow)
• Download retry logic handles HTTP 403 with clear rate-limit message
• File locking prevents concurrent download races
• hyperframes tts "Hello" -o test.wav works end-to-end
• Verify TTS failure rate drops after deploy