fix(speaker): cross-condition negative-exemplar veto via condition transport

[r3dbars]

What

Strengthens the negative-exemplar veto (#1487) so it fires cross-condition, implementing recommendation #3 of the PR #1493 eval (docs/speaker-eval-exemplar-delta-2026-07.md): "give a profile's negative exemplars the same multi-condition treatment as its positive ones."

Why

PR #1493 measured the veto as "sound but regime-limited": after a correction it removes 46% of repeat wrong-matches in-room (AMI) but only 12% cross-condition (VoxCeleb). A rejected in-room sample and a later telephone/VoIP sample of the same rejected impostor sit too far apart in embedding space for the raw cosine to clear the 0.80 veto floor — so a correction made in one audio condition failed to protect the owner in another. The positive side already closed this exact gap with multi-exemplar voiceprints (#1488); the negative side (a single rejected embedding) had no analog.

How

SpeakerNegativeExemplarPolicy gains a condition-transported negative similarity. Beyond raw cosine to each stored rejected sample, the candidate is compared against each rejected sample transported along the profile's own observed condition shifts — unit(exemplar) − average, reusing the positive multi-exemplar voiceprints. Channel/condition shifts (clean → compressed remote → telephone) are largely speaker-independent, so a rejected sample shifted by a condition the profile has actually seen approximates the same rejected voice returning in that condition.

Owner-safe by construction (the #1493 guardrail is the priority):

• Transport only along +(exemplar − average) — real, forward, profile-observed conditions. Bidirectional/synthetic directions were measured to leak owner-collateral and are not used.
• A profile with no positive exemplars derives nothing → reduces exactly to the raw maxNegativeSimilarity → single-condition profiles (and any store with no negatives) are byte-for-byte unchanged.
• The 0.80 floor and the ≥ positiveSimilarity owner gate are untouched. Transport only widens which impostor returns are caught; it never lowers the floor, so a genuine owner — closer to its own fingerprint than to a transported impostor sample — is still protected.

Results — real qmatrix corpus (SpeakerExemplarDeltaEvalTests)

Same harness/corpus/operating point as #1493. RAW = single stored rejected sample (pre-fix); TX = condition-transported (this PR); both use the unchanged 0.80/≥pos gate.

AMI (in-room, 147 profiles, 535 cross-condition wrong re-matches, 5773 owner checks)

metric	RAW	TX	Δ
cross-condition vetoed-among-re-match	0.350	0.374	+7.0% rel
aggregate vetoed-among-re-match	0.460	0.483	+0.023
owner-collateral	2.17%	2.32%	+0.16pp (+9 of 5773)

VoxCeleb (single-utterance, 30 profiles) — byte-identical (0.123→0.123; owner-collateral 1.56%→1.56%): no condition structure to transport along, and no regression.

On VoxCeleb's 12%

It doesn't move — and that is correct, not a miss. VoxCeleb is single-short-utterance-per-meeting, so a "cross-condition" replay differs from the rejected sample in content/session as much as in channel; a channel transport can't (and shouldn't) bridge that. Real returning Transcripted users are AMI-like (repeated appearances across conditions → multi-exemplar structure), which is the regime this improves. Pushing VoxCeleb's 12% harder was investigated across floor-lowering, ramped owner-margins, global channel subspaces, and orthogonal-complement distance — every mechanism trades owner-collateral for veto coverage at a ≈1:1-or-worse absolute rate (best case removed +39 wrong re-matches while adding +56 owner false-vetoes), which is net-negative in production and violates the #1493 guardrail. Condition transport is the one lever that gains cross-condition coverage without a net owner cost.

Full write-up: docs/speaker-eval-negative-veto-cross-condition-2026-07.md.

Tests

• SpeakerNegativeExemplarPolicyTests 12/12 — new fixtures: same-impostor-different-condition (veto now fires), owner-different-condition (no false veto), no-exemplars-equals-raw (byte-identical fallback).
• SpeakerNegativeExemplarTests 7/7, SpeakerNamingSimulationRunnerTests 7/7, full Core suite 714 passed / 0 failed (13 corpus-gated skips), SpeakerExemplarDeltaEvalTests passes on the real corpus.

Coordination

Orthogonal to #1496 (average-based auto-accept margin): that PR changes the auto-accept margin computation; this changes the negative-veto representation. They touch SpeakerMatchingService.swift and SpeakerExemplarDeltaEvalTests.swift in different regions. #1496 should land first; I'll rebase this onto it, keeping both changes.

Refs #1493, #1487, #1488.

🤖 Generated with Claude Code