docs(adr): ADR 0116 revision V7 — component-scoped weighted HAC with highest-valid-cut #1316
Loading…
Reference in a new issue
No description provided.
Delete branch "docs/adr-0116-highest-valid-cut"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Summary
Amends ADR 0116 with decision V7 (revision, 2026-10-19): the first live v4 run (job
fc05c137, ~31 min at pool 21194) showed that medoid nodes bridge previously separate U×U components into giant U∪R instances — and since HAC cost is quadratic in instance size ((a+b)²/(a²+b²) ≤ 2per bridge), the global-HAC wording is unworkable at production scale.V7: keep v4's identity-aware weighted HAC, but scoped per U×U component (the fine decomposition v3 uses) with only the representatives adjacent to it (rᵢ ≪ nᵢ), and replace the fixed cut with the highest-valid-cut: h* = the highest merge height ≤ clustering_max_distance at which every cut subset carries ≤ 1 identity bit (backward linear scan with bit-OR tracking, O(k)); if nothing ≤ threshold is valid, the deepest valid height anywhere. Mixed remainders surface via V6 margin demotion. Runtime returns to v3-class (Θ(d·Σ(nᵢ+rᵢ)²) ≈ Θ(d·Σ nᵢ²)) — honestly stated as NO speedup vs v3; the gain is quality (mixed components split into correct per-identity cards instead of ambiguous discard).
Also amends the validation plan: new G-A′ gate (CV re-run for the V7 lane; ambiguous target ≤ 100; same-person splits benign majority; harness runtime ≤ 2× v3 lane) + G-B fixtures 9–11 (cut unit tests, component-scoped equivalence incl. the rep-touching-two-components counter-case, runtime regression pin). Former V7 (Rollout) renumbered V8.
ISSUE-169/PR #1315 (compatiblePersonIds) is part of this baseline (merged before the revision).
Test plan
python3 -m json.tool); programmatic assertions over status/date/title/links/decision blocks; word-diff reviewed.