Add Transcripted Lab experiment workbench and CLI

[r3dbars]

What this adds

A separate local Transcripted Lab product for running repeatable experiments against Transcripted's real production and repository-owned test paths.

• Native SwiftUI Transcripted Lab.app
• Shared TranscriptedLabKit experiment/report engine
• Headless transcripted-lab CLI backed by the same contract
• Versioned JSON reports, run history, and baseline/candidate comparisons
• Hard gates that fail before the numeric score can hide a serious regression

Benches

• Runtime Snapshot — transcription elapsed/RTF, dictation start, first sample, stop-to-paste/done, fallback/retry events
• Dictation Bench — production/native/pre-resampled/chunked variants, compute units, finalization order, repetitions, Auto Enter, silence guardrail, stable output hashes
• Transcription Bench — existing local corpus validation and Transcripted-vs-truth comparison
• Speaker Bench — labeled-audio threshold sweeps plus the existing frozen ASK/SUGGEST/AUTO auto-research path
• QA Bench — quick/deep/full/UI/packaged/artifact/synthetic/live repository lanes

Safety rules

The score cannot average away:

• cross-person speaker false merges
• missing final dictation text
• silence producing text
• identical input producing unstable output hashes
• blocking QA failures
• timeouts or non-zero experiment exits

Local verification completed

• swift test --package-path Tools/TranscriptedLab --jobs 4 — 5/5 focused tests passed
• swift build --package-path Tools/TranscriptedLab --product transcripted-lab
• CLI help and command parsing exercised
• app-bundle script syntax validated

The new macOS workflow builds and tests the Foundation kit/CLI, compiles the SwiftUI app, and verifies the ad-hoc-signed app bundle.

Current intentional limits

• Dictation Bench currently measures speed, delivery integrity, silence behavior, and repeatability. A dedicated labeled dictation WER/CER corpus is the next accuracy lane.
• Transcription corpus accuracy and production runtime speed are separate lanes today.
• Speaker auto-research keeps its native multi-gate holdout report rather than flattening false automatic names and contamination checks into a misleading single score.

[r3dbars]

Verification complete

All automated gates are green on 8f15c626:

• Transcripted Lab focused tests: 5 passed
• transcripted-lab CLI build: passed
• SwiftUI TranscriptedLab build: passed
• ad-hoc-signed Transcripted Lab.app bundle verification: passed
• repo hygiene: passed
• Transcripted app build + premium performance budget: passed
• fast tests + E2E smoke: passed
• Core package + integration + all Tools package tests: passed

First local launch after checking out this branch:

Tools/TranscriptedLab/script/build_and_run.sh

CLI entrypoint:

swift run --package-path Tools/TranscriptedLab transcripted-lab help

I am intentionally keeping the PR in draft while the next likely slice remains explicit: a labeled dictation/transcription WER/CER corpus, normalized speaker auto-research holdout metrics in the UI, and bounded multi-arm campaign recipes similar to Tilde Lab.