[nightly-review] Fix captured input attenuation diagnostics

[r3dbars]

Summary

• Capture the selected meeting input device scalar at recording start.
• Make issue Mic audio recording is a lot quiter #500 attenuation classification prefer captured-device volume drops over default-input drops.
• Guard captured-volume deltas so device recovery cannot compare unrelated input devices.

Nightly Signals Reviewed

• No open [nightly-review] or [nightly-regression] PRs were found before editing.
• origin/main last 24h included Classify meeting mic attenuation for issue #500 #901, [nightly-review] Force recovery after ready dictation start failures #903, and Surface stale Claude helper repairs #904.
• Sentry last 24h: APPLE-MACOS-14 2x 1.1.44 dictation start timeouts, APPLE-MACOS-1B 1x 1.1.44 capture degraded, APPLE-MACOS-9 1x old 1.0.6 app hang.
• PostHog last 24h: 49 app launches on 1.1.44, 40 meeting transcripts saved, 2 dictation start failures on Bluetooth output route, 1 meeting degraded, 1 no-speech skip.

Top Risks

1. Issue Mic audio recording is a lot quiter #500 attenuation instrumentation could misclassify the captured input route; fixed here.
2. Dictation Bluetooth ready-start timeout still appears in Sentry/PostHog, but latest events line up with the pre/[nightly-review] Force recovery after ready dictation start failures #903 window, so watch next run.
3. Update check check_timed_out/unknown is noisy in PostHog; not patching tonight because it needs Sparkle/runtime evidence.

Verification

• bash scripts/dev/agent-preflight.sh
• git diff --check
• TRANSCRIPTED_DISABLE_FILE_LOGGER=1 bash build-deps.sh --force
• TRANSCRIPTED_DISABLE_FILE_LOGGER=1 bash build.sh --no-open
• TRANSCRIPTED_DISABLE_FILE_LOGGER=1 bash run-tests.sh
• TRANSCRIPTED_DISABLE_FILE_LOGGER=1 bash run-integration-smoke.sh
• TRANSCRIPTED_DISABLE_FILE_LOGGER=1 swift test
• /Users/redbars/.codex/skills/codex-review/scripts/codex-review --mode branch

[r3dbars]

verdict: no-findings-but-human

findings: none. The captured-input scalar now owns the attenuation classification when present, default-input facts remain reported separately, and the added tests cover redirected, flat, and unreadable captured-input cases.

actions: no fix pushed.
current_head_sha: 2252702
fix_commit: none
checks: repo-hygiene pass; gh pr view; gh pr diff --patch; gh pr checks; git diff --check origin/main...origin/codex/nightly-review-captured-input-volume-20260530; Meeting/Core/Observability/Test docs reviewed.

manual_merge_handoff:
head_sha: 2252702
verdict: no-findings-but-human
blockers: draft PR; human signoff for meeting audio diagnostics and telemetry surface
checks: repo-hygiene pass; PR metadata/diff/checks reviewed; diff-check clean; Meeting/Core/Observability docs reviewed