Refresh docs to match what shipped

[bradburch]

Brings CLAUDE.md, README.md, and docs/local-llm.md in line with #18, #19, and #20.

Wrong, not just stale

Three CLAUDE.md claims would have actively misled a future session:

Claim	Reality
"CoachingLLM has two implementations"	Three — and only the Anthropic one can enforce a JSON schema; the other two share one prose contract (formatAppendix states the rules, decodeCoaching enforces them)
System audio via ScreenCaptureKit	A CoreAudio process tap. SCK is precisely what recorded an entire Continuity call as digital silence
Deployment floor 14.0	14.2 since CoreAudio taps, in Package.swift and LSMinimumSystemVersion

The capture section now also records the two tap invariants that are easy to undo by accident: the aggregate device deliberately has no output sub-device, and padSilenceToNow exists because a tap delivers nothing while output is idle — remove it and every "them" segment after a gap lands early in the merge.

What actually cost time this session

• The documented AppleScript for opening the main window silently no-ops without a preceding activate — it reports success and leaves the window count at 0. The snippet in CLAUDE.md was incomplete.
• entire contents doesn't just return empty on the Settings pane, it errors (-1700). Counting the Form's section groups works as a structural probe. screencapture is not a fallback — the shell lacks Screen Recording.
• Comparing mic vs sys RMS across the recordings directory is the fastest capture diagnostic, and it separates "capture is broken" from "this call type is excluded" in one pass. audioop was removed in Python 3.13, so the obvious one-liner doesn't work.
• Never run make-app.sh while a recording is live — it overwrites the running executable.
• A Finder-launched .app inherits a minimal PATH, so spawning a helper binary needs absolute-path search; /usr/bin/env resolves in a shell and fails in the bundle.

Added a third entry to "green tests prove less here than usual": capture that runs, writes correctly-sized files, and records silence. Every mocked recorder test passed while SCK wrote zeros.

One new rule, the converse of an existing one

CLAUDE.md already says don't split what shares a contract. This adds don't narrow what already serves two callers — that bug shipped twice this session: a shared query filtered for one of its two callers (dropping mock interviews from export), and a second reader added to one file descriptor (racing the streaming drain).

User-facing

README and docs/local-llm.md describe all three providers. The subscription path's tradeoffs — ~15–20k overhead tokens per debrief, rate limits, unsupported integration — are stated in local-llm.md, since that's the page someone lands on when they're trying to avoid API cost and comparing options.

201/201 unit tests pass (docs-only change; run as a sanity check).

🤖 Generated with Claude Code