Add Deepgram backend: transcribe audio over Whisper's 25 MB limit (45 min+ videos)

[drlee91]

Problem

Groq and OpenAI Whisper both reject uploads over 25 MB with HTTP 413. At the 64 kbps mono extraction rate that's roughly 45 minutes of audio — so any longer caption-less video currently comes back frames-only, with no transcript possible at all:

[watch] audio: 27625 kB — uploading to groq Whisper…
[watch] whisper fallback failed: Whisper request failed: HTTP Error 413: Payload Too Large
       — {"error":{"message":"Request Entity Too Large",...,"code":"request_too_large"}}

This bites in practice: long tutorials, podcasts, talks, VODs — and YouTube caption pulls can also fail with 429 rate limits, which makes the audio fallback the only path for those videos.

Approach

Add Deepgram's pre-recorded API (nova-2) as a third backend. Its file limit is ~2 GB, so the 25 MB cliff disappears. Routing keeps current behavior unchanged for existing users:

• Audio ≤ 24 MB: exactly as before — Groq preferred, OpenAI fallback. Deepgram additionally catches the case where Whisper errors out.
• Audio > 24 MB: routed straight to Deepgram instead of failing with 413. Without a Deepgram key, Whisper is still attempted so the user sees the clear 413 error.
• Standalone: if DEEPGRAM_API_KEY is the only configured key, Deepgram serves as the primary backend.
• Override: --whisper deepgram forces it.

Design notes

• Pure stdlib, matching whisper.py — urllib only, no SDK dependency. Same {start, end, text} segment shape (from Deepgram utterances, with paragraph/transcript fallbacks), same SystemExit error contract, same key resolution (env → ~/.config/watch/.env → ./.env).
• detect_language=true keeps it language-agnostic like the Whisper path.
• The module is named deepgram_backend (not deepgram) so it can never shadow or be shadowed by the real Deepgram SDK.
• setup.py accepts DEEPGRAM_API_KEY as a valid transcription key, scaffolds a commented placeholder, and includes it in the install hints. --check/--json semantics are unchanged otherwise.
• Considered chunked uploads to stay within Whisper's 25 MB instead, but that needs segment-boundary handling, per-chunk timestamp offsetting, and N sequential uploads — a second provider that follows the existing backend contract is simpler and also gives users a provider choice. Happy to adjust if you'd prefer the chunking route.

Verification

• 59-min YouTube video (caption pull 429'd, audio 27 MB): routed to Deepgram, returned 628 timestamped segments, full report, exit 0. Previously: frames-only.
• Small file regression: 40 s local clip without flags → still transcribed via Groq (unchanged default).
• Forced: --whisper deepgram with a Groq key present → Deepgram used.
• Standalone: only DEEPGRAM_API_KEY configured → Deepgram auto-selected as primary; setup.py --json reports ready.
• No keys at all: updated guidance message lists all three options; --check exits 3 as before.

Docs (README + SKILL.md) updated accordingly.