Speaker-naming simulator + same-voice consolidation to cut speakers-to-name

[r3dbars]

What & why

Speaker naming feedback says the review sheet can ask the user to name 4-7 "people" after a 1-on-1 call. The likely failure is offline VBx over-segmenting one remote voice into several large clusters that survive small-cluster absorption.

This PR does two things:

• adds a conservative same-voice consolidation pass in EmbeddingClusterer
• adds a deterministic speaker-naming simulator that measures the user-facing review-row count, not just raw diarizer cluster count

Changes

Same-voice consolidation

• EmbeddingClusterer.postProcess now runs consolidateSameVoiceClusters after small-cluster absorption and before DB-informed split.
• Offline system and opt-in mic diarization paths both get the pass through their existing postProcess(... pairwiseMergeThreshold: nil) calls.
• The pass agglomeratively merges clusters above the SpeakerNamingPolicy auto-accept bar (> 0.88) and recomputes centroids after each merge.
• Added a boundary test so clusters at exactly 0.88 similarity stay separate, matching the naming policy edge.

Speaker-naming simulator

scripts/ops/speaker-naming-simulator.py now reports:

• review_before / review_after
• expected review rows
• expected labels
• remote vs local role
• false-merge flags

Coverage now includes:

• cold unknown 1:1 over-segmentation: 5 review rows -> 1
• named repeat speaker persistence: saved name auto-applies, 0 review rows
• tentative known speaker: 1 confirmation row, not duplicate rows
• remote small/crowded groups
• near-threshold similar distinct voices
• local mic default-off behavior: You, 0 review rows
• opt-in local split behavior: room speakers become local review rows

Docs

• docs/qa-test-bench.md documents the simulator as review-row/label/false-merge coverage.

Local verification

Passed on macOS in this worktree:

python3 -m py_compile scripts/ops/speaker-naming-simulator.py
scripts/ops/speaker-naming-simulator.py
scripts/ops/speaker-naming-simulator.py --sweep
git diff --check
bash scripts/dev/agent-preflight.sh
bash -n scripts/ops/transcripted-qa-bench.sh
bash -n scripts/ops/run-local-summary-fixture.sh
python3 -m py_compile scripts/ops/validate-meeting-corpus.py
python3 -m py_compile scripts/ops/compare-meeting-corpus.py
bash build-deps.sh --force
bash build.sh --no-open
bash run-tests.sh
bash run-integration-smoke.sh
swift test
bash scripts/ops/run-local-summary-fixture.sh
swift test --package-path Tools/TranscriptedQA
bash scripts/ops/transcripted-qa-bench.sh --mode quick

Notable counts:

• run-tests.sh: 5255 passed, 0 failed
• swift test: 474 passed, 1 skipped, 0 failed
• Tools/TranscriptedQA: 34 passed, 0 failed
• quick QA bench: PASS, 6/6 checks

Review status / blockers

Keep this PR draft for now.

• codex review --base origin/main was attempted but hit the Codex account usage limit before producing findings.
• Claude Code second-opinion review was attempted through claude -p but hung without output and was interrupted.
• CI should re-run on the pushed head 80ca3e7f.
• No live audio/manual corpus proof has been run; this simulator is deterministic model coverage, not real speaker-ID accuracy proof on Justin's recordings.

Merge recommendation

Hold as draft until an independent review completes cleanly and, ideally, at least one real or corpus meeting verifies that the reduced review-row count does not hide a true distinct speaker.

[r3dbars]

Rebased onto main + cleared the merge-state CI failure

This branch was stale (forked at 0c72c29c, before SpeakerNamingSimulationRunner landed on main). On the merged tree CI was red on a single test:
SpeakerNamingSimulationRunnerTests.testSimulationReportFlagsConfusionFalseMergeAndFalseSplit (XCTAssertFalse at :107 + XCTAssertTrue at :110).

Root cause (verified locally, not just assumed)

The negative-control suite built its false split from two clusters with identical alex embeddings. The new same-voice consolidation pass (cosine > 0.88) correctly merges identical voices, so the two clusters collapsed into one review row and the false split disappeared — falseSplitIndicators came back empty. Reproduced on the rebased tree; the report showed False split indicators: none while false-merge/confusion detection still fired. The detector was never broken — consolidation (the feature) legitimately removed the split the test depended on.

Fixes

1. Rebased the 2 PR commits onto current main (clean, no conflicts; main had not touched any PR file since the fork point).
2. Negative control now models a drifted same-voice over-segmentation (near(alex, degrees: 35) ≈ 0.82 cosine, below the 0.88 bar) instead of an identical voice. Consolidation correctly leaves it split, the user mislabels it, and the runner's unchanged detector flags a genuine false split. No assertion was weakened — the scenario now exercises the residual false-split risk that consolidation cannot fix.
3. Threshold drift guard: named both thresholds (SpeakerNamingPolicy.autoAcceptSimilarityThreshold and EmbeddingClusterer.sameVoiceConsolidationThreshold), wired the consolidation defaults to the constant, and added EmbeddingClustererTests.testConsolidationThresholdMatchesAutoAcceptBar asserting they stay equal. Pure refactor — both remain 0.88.

Tests (local swift test, the same path CI runs)

• Full package: 486 executed, 1 skipped, 0 failures (was 1 failure pre-fix).
• Speaker/clusterer/pipeline suites: 128 passed, 0 failed — incl. EmbeddingClustererTests (12, with the new guard) and SpeakerNamingSimulationRunnerTests (6, with the previously-failing test now green for the right reason).

Still manual / unproven (unchanged by this PR)

No real-audio or corpus validation was run — speaker attribution on real meetings stays manual and unknown. The honest claim remains guardrails against over-merging, not solved speaker attribution: aggressive same-voice consolidation can still over-merge genuinely distinct speakers (e.g. a 6-person call collapsing). Recommend a real-corpus pass before relying on the review-row reduction in production.

[r3dbars]

Codex review follow-up: known-profile consolidation guard

Pushed f77aa172 to keep same-voice consolidation from merging two plausible different known speakers before review. The fix adds a lower known-profile conflict guard (aligned to the 0.70 DB match floor) plus regressions for both exact known-profile conflicts and looser profile matches below the 0.88 consolidation bar.

Local proof on the final patch:

• bash build-deps.sh --force
• bash build.sh --no-open
• bash run-tests.sh (5350/5350)
• bash run-integration-smoke.sh
• swift test (488 executed, 1 skipped)
• swift test --filter EmbeddingClustererTests (14/14)
• swift test --filter SpeakerNamingSimulationRunnerTests (6/6)
• scripts/ops/speaker-naming-simulator.py + --sweep
• bash scripts/ops/transcripted-qa-bench.sh --mode quick (6/6)
• codex-review --mode branch: clean, no accepted/actionable findings

Hold reason: PR is still draft; GitHub repo-hygiene passed after push, but build-and-test was still pending when I stopped watching. Real-audio/corpus validation is still not proven, so keep this YELLOW until CI is green and a real meeting corpus/manual audio pass confirms attribution behavior.