check_pr_ont_delta: §11.1 names docs/specifications/** whole, so every design spec reads as a sweep — three maintainer PRs red for a delta that does not exist; narrow the sink to the release spec, spec + guard together, with a predicate self-test #3535

Measured 2026-09-19 (aprender-ba): three maintainer PRs — #3499 (a guard-regex fix), #3420 (PP-QUANT-001, a design spec), #3424 (PP-ARCH-001, a design spec) — are RED on guard-tree with FAIL no ont-delta: line. §11.1: a sweep PR closes with a delta. None of them is a sweep: each touches exactly one file under docs/specifications/ and nothing else the rule names.

scripts/check_pr_ont_delta.sh is FAITHFUL to APR-RELEASE-001 §11.1 as written — the sentence names docs/specifications/** whole as a prose sink — so the predicate cannot be narrowed in the script alone; the sentence and the guard move together. A design spec (a *-MASTER.md, a new feature's specification) is the finding's HOME, not a sink a sweep writes into; the sink under docs/specifications/ is the train's own spec, APR-RELEASE-001, where sweep findings are recorded.

Second finding: the guard's self-test had no predicate leg — all 15 rows feed check_body_text, so sweep_reason could classify every PR either way and stay green. The patch (aprender-ba, 55 lines) adds predicate_self_test (10 rows) with the falsifying row "a DESIGN spec is not a sweep": under today's sentence it fails (got 0, want 1), under the amended sentence + sink list it passes.

Fix: §11.1's sentence names docs/specifications/APR-RELEASE-001-*.md as the sink (not the directory); ONT_SWEEP_SINKS follows; the predicate self-test lands with the design-spec row at want=1 constant. Interim for the three PRs: an ont-delta: none <reason> body line.