[abrichr]

Builds the map row "Use spoken narration while recording to learn why a branch was taken."

Stacked on #1276 (and #1260 through it). Merge those first; this diff then shrinks to the narration work. The install line uses the audio extra from #1271.

What it does

• Capture scrubs the transcript with openadapt-privacy before storing it, and refuses --audio before the microphone opens when the scrubber is missing. Scrubbed sentences go to narration.json, timed on the recording clock. Per-word text is no longer stored. capture transcribe scrubs too.
• Flow carries the sentences into the recording and lines each one up with a step (narration.json). openadapt-flow narration attach does the same for any recording from a WAV or a transcript, so web, native, RDP and Citrix all work the same way.
• induce reads a sentence near a branch ("If the insurance is Medicare, skip the copay screen") and proposes a guard on a recorded field (param_equals) or on screen text (text_present). Works for optional steps and for two-path divergences.
• Certification. A proposal becomes a guard only when every recording agrees, --probe-on confirms it (new narrated reading), or a person accepts it (--accept-narrated LOC --accepted-by NAME). A contradicting recording keeps the point refused, and a person can't accept it.
• teach-branch uses a sentence to pick the narrowing parameter when every example and normal run agrees.
• Explain back. Certified guards keep their sentence in Workflow.guard_sources (omitted when empty, so existing bundles are byte-identical). visualize (HTML, Mermaid, JSON) and repair show show it. Non-local visualize profiles drop the sentence.
• Model. --narration-model is opt-in, compile time only, on the customer's key, scrubbed outbound, and its answer is validated. Runs make no model calls.

Tests

Synthetic data only. packages/capture/tests/fixtures/narration/medicare_copay.wav is speech generated locally with macOS say (invented name).

• narrated branch becomes a proposed guard; certifies when recordings agree; a contradicting recording keeps it refused; transcript scrubbed (capture and flow); healthy runs make no model calls
• probe confirms a guard the recordings can't check; person acceptance; divergence branch; every backend; visualize; teach-branch; CLI
• End to end with real local Whisper (tiny.en) on the fixture: passes locally; skipped in CI unless the model is cached (never downloads)

Local: tests/test_narration_rules.py 29 passed, test_teach_branch_narration.py 4, capture test_narration.py 7 + test_audio_privacy.py 15; a 37-file Flow slice (induction, probe, teach, visualize, repair, privacy, capture adapter) 690 passed.

Docs: packages/flow/docs/NARRATION.md, capture README/COMPLIANCE/DESIGN, root CAPABILITIES.md row (build_llms_full.py re-run).

🤖 Generated with Claude Code

[abrichr]

Included in batch #1384, which runs CI once for several reviewed PRs. This PR is marked merged when #1384 merges. Please don't push here unless asked; a new commit leaves it out of the batch.