Exempt faithful preserve-by-default splits from the pre-merge gate

Narrows the locked 'full regression re-run before merge' rule
(RUBRIC_AUDIT.md / REGRESSION_GATE.md). A faithful preserve-by-default
split — structural c2-split that only redistributes a locked
old-c2=T into c2=T+c3=T, with zero realignments — is exempt: it
changes no locked value's meaning, and since we don't act on grader
output, a gate run can't change any expected value (it would only
produce an informational Sonnet-calibration snapshot). Local
validators + the faithful decomposition suffice.

Gate stays REQUIRED for any change that realigns expected values or
any new authoring. Recorded to avoid silent process drift (the kind
that bit PR #146).

Author: claude

docs/REGRESSION_GATE.md

@@ -27,6 +27,15 @@ This capability asymmetry has a hard consequence for gate triage:
 
 This principle was added 2026-05-21 after the c2-split batch surfaced the asymmetry: several early gate realignments flipped Opus-authored values to match Sonnet's stricter/looser reads, which was backwards.
 
+## When the pre-merge gate is required vs. optional
+
+`RUBRIC_AUDIT.md` § Decisions locked originally required a full pre-merge gate run for *every* rubric change. The grader/gold-standard principle above narrows that:
+
+- **Gate REQUIRED** for any change that *realigns* expected values, or any new authoring — i.e. anything where the grader's output could change a recorded value or where there's no locked value to fall back on.
+- **Gate OPTIONAL (informational only)** for a *faithful preserve-by-default split* — a structural c2-split that only redistributes a locked `old-c2=T` into `c2=T` + `c3=T` (and `old-c2=F` into a fair-read name/mechanism judgment), with **zero realignments**. These don't change any locked value's meaning, and since we don't act on grader output, a gate run can't change a single expected value — it would only produce a documented snapshot of Sonnet calibration gaps. Local validators (lint, `ingest_units --check`, `run_regression_set --check`, pytest) plus the faithful decomposition are sufficient to merge.
+
+Rationale: the original "always gate" rule predated this principle; its purpose was to catch grader miscalibration we'd *act on*. When we preserve by default, that purpose doesn't apply. Running the gate remains welcome as an informational snapshot, just not a merge gate. (Recorded 2026-05-21, during the MEDIUM-batch rollout, to avoid silent process drift.)
+
 ## Procedure (operator)
 
 For any PR that edits `content/units/**` (rubric text) or `content/regression-sets/**`:

docs/RUBRIC_AUDIT.md

@@ -8,7 +8,7 @@
 ## Decisions locked (2026-05-20)
 
 - **Approach:** SPLIT. Every flagged criterion in this audit is split into two criteria. Rubrics for Units 2–9 grow from 3 → 4 criteria. Units 10–13 use SPLIT as well (Option A — new criterion) rather than the Option B "move meta-claim into existing c3" path, for cross-unit consistency.
-- **Regression backfill:** every unit's regression set is re-run end-to-end before its PR merges. No spot-checks. Operator procedure in `docs/REGRESSION_GATE.md`.
+- **Regression backfill:** every unit's regression set is re-run end-to-end before its PR merges. No spot-checks. Operator procedure in `docs/REGRESSION_GATE.md`. *(Narrowed 2026-05-21: a faithful preserve-by-default split with zero realignments is exempt from the pre-merge gate — see `docs/REGRESSION_GATE.md` § When the pre-merge gate is required vs. optional. The gate stays required for any change that realigns expected values.)*
 - **PR cadence:** one PR per unit (pilot first, then batched if the pilot is clean). Each PR includes the rubric edit, `ingest_units` re-run, regression-set updates, and a regression-run report.
 - **C1 / c3 follow-up:** deferred. Separate audit doc after this sweep lands, not folded into per-unit PRs.