Add --llm-reduce: one LLM prompt over all batch-transcribe results

[alexkroman]

Summary

Adds a repeatable --llm-reduce 'PROMPT' flag to assembly transcribe — a map-reduce for batch transcription. --llm is the per-source map (runs over each transcript); --llm-reduce is the reduce (one LLM-Gateway call over all sources' results).

assembly transcribe --from-stdin --concurrency 3 --speaker-labels \
  --llm 'Judge diarization quality; output JSON {speaker_count, issues, score}' \
  --llm-reduce 'Rank these videos worst-to-best and summarize the failure modes'

Behavior

• Batch mode: after every source transcribes, each source contributes its last --llm output (or its transcript text if no --llm ran), concatenated under ### Source: headers, and one reduce chain runs over the combined text. The result prints to stdout; the progress table is routed to stderr so stdout stays pipe-clean. Under --json, a final additive {"type":"reduce","model","prompts","output"} NDJSON record follows the per-source result records.
• Single-source mode: nothing to aggregate, so the reduce prompts extend the --llm chain over the one transcript.
• A reduce only runs after the batch fully succeeds (_summarize raises on any failure first), and is skipped — with a stderr warning — when there is nothing to reduce (avoids a billable call over empty input).

Touches only transcribe's own modules (commands/transcribe.py, app/transcribe/{run,batch}.py) plus docs/snapshot — no shared-file edits.

Test Plan

• ./scripts/check.sh green: lint, mypy/pyright, 100% patch coverage, diff-scoped mutation gate, build + twine
• New tests/test_transcribe_reduce.py: flag plumbing, single-source chain, batch reduce (map-output + transcript-text fallback), {"type":"reduce"} NDJSON, stdout/stderr routing, empty-reduce skip, helper edge branches
• transcribe --help snapshot regenerated (incl. the t alias)
• REFERENCE.md + README documented; docs-consistency gate passes
• Final code review addressed (empty-combined-text guard added)

Design spec and implementation plan: docs/superpowers/specs/2026-06-16-llm-reduce-batch-transcribe-design.md, docs/superpowers/plans/2026-06-16-llm-reduce.md.

🤖 Generated with Claude Code

[aikido-pr-checks]

The llm.run_chain call receives transcript_text assembled from user transcripts; do not send unsanitized user transcripts to external services without sanitization or explicit opt-in.

Details

✨ AI Reasoning
​The new _run_reduce calls llm.run_chain(api_key, transform.reduce_prompts, transcript_text=combined, ...). The combined string is constructed from user transcripts (see _gather_reduce_inputs) and is passed unchanged as transcript_text to an external LLM-Gateway. This is effectively logging/exfiltrating user-controlled content to a third party and may leak PII or otherwise sensitive data if not sanitized or consented to.

🔧 How do I fix it?
Keep sensitive data such as emails, passwords, and tokens out of logs. When logging values tied to a user, prefer a safe identifier like a user ID over the raw input, and strip line breaks from any user-provided text you do log.

_{Reply @AikidoSec feedback: [FEEDBACK] to get better review comments in the future.
Reply @AikidoSec ignore: [REASON] to ignore this issue.
More info}