[bug-hunter] Make speaker profile merges atomic

[r3dbars]

Summary

• move name transfer, target update, provenance, exemplar cleanup, and source deletion into one throwing SQLite transaction
• propagate prepare, step, row-count, begin, and commit failures to callers
• stop post-merge cleanup when Settings observes a failed merge
• add deterministic rollback coverage for transaction and statement failures

Proof

• bash build-deps.sh --force
• bash build.sh --no-open
• bash run-integration-smoke.sh
• focused merger tests (13 passed)
• speaker tests (254 passed, 12 external-fixture skips)
• full swift test (768 passed, 13 external-fixture skips)
• bash run-tests.sh (13,189 passed)
• git diff --check

Manual boundary

The Settings speaker-merge flow was not exercised manually.

[r3dbars]

Repair pushed at 25789b6d0bcd6be53a6c9a639ff5ed6cd3405c89.

Review fixes:

• database merge now commits before clip promotion/deletion, so a rollback leaves both clip sets untouched
• missing/incompatible profiles now throw instead of being treated as a successful no-op
• naming updates preserve mutation order inside one database transaction and restore the transcript if persistence fails
• duplicate and same-name sweeps continue safely after a failed pair

Proof:

• TRANSCRIPTED_DISABLE_FILE_LOGGER=1 bash run-tests.sh: 13,192/13,192 passed
• focused side-effect regression: 2/2 passed
• build source-list validation, shell syntax, and git diff --check: passed
• final codex review --base origin/main: no actionable correctness issues
• independent risk review: /Users/redbars/.codex/maestro/runs/20260717T114927Z-pr-1546-review-claude
• targeted Swift package tests are locally blocked in this fresh worktree by the pre-existing missing FluidAudio dependency; GitHub CI is the package/build authority for this new head

Keeping this PR draft because the Settings speaker-merge UI/manual gate is still unproven.

[r3dbars]

Fresh SwiftPM CI failure repaired at 658626ab.

What failed:

• spm-tests could not compile SpeakerProfileMerger.swift because the new local closure referenced instance members without an explicit self capture.
• A fresh local package compile then exposed the new rollback test comparing SpeakerProfile values even though the type is not Equatable.

Fix:

• Added the behavior-neutral [self] closure capture.
• Made the rollback assertion compare every persisted SpeakerProfile field explicitly, without widening the production type API.

Proof on exact head before push:

• bash build-deps.sh --force — passed; full dependency graph and TranscriptedCore compiled.
• swift test --filter SpeakerProfileMergerTests — 16/16 passed.
• bash run-integration-smoke.sh — core smoke + wake smoke passed; recovery merge package tests 7/7 passed.
• bash run-tests.sh --filter SpeakerProfileMergeSideEffectCoordinator — 2/2 passed.
• git diff --check — passed.
• codex review --uncommitted — no actionable findings; independent Claude diff review also returned no findings (MAESTRO_PROOF=/Users/redbars/.codex/maestro/runs/20260717T121625Z-pr1546-review-claude).

Keeping this PR draft: the Settings speaker-merge UI path still needs the stated manual proof before it is merge-ready.

[r3dbars]

Disposition: superseded by current-main speaker merge replacement #1571, merged as e1f7440. The original branch was based on stale main and is not being merged.