objectstack-ai
diff --git a/‎AGENTS.md‎
Lines changed: 24 additions & 0 deletions b/‎AGENTS.md‎
Lines changed: 24 additions & 0 deletions
diff --git a/‎docs/audits/2026-09-self-test-shape-census.md‎
Lines changed: 19 additions & 0 deletions b/‎docs/audits/2026-09-self-test-shape-census.md‎
Lines changed: 19 additions & 0 deletions
diff --git a/‎scripts/pm/check-skill-line-ratchet.mjs‎
Lines changed: 20 additions & 1 deletion b/‎scripts/pm/check-skill-line-ratchet.mjs‎
Lines changed: 20 additions & 1 deletion
@@ -1026,6 +1026,30 @@ registry? Add it to `OPEN_CAPABILITY_REGISTRIES` in the same PR that fixes it.
 
 ---
 
+## Writing a `--self-test` — it must be capable of failing when it runs nothing
+
+**Floor — pin battery NAMES, never one total.** Declare a frozen roster of battery name → minimum
+case count, register every case against a battery, and fail when a declared battery registers fewer
+cases than its pin, when a registered case names no declared battery, or when the roster itself
+falls below its pinned battery count. ⛔ A printed case count is EVIDENCE, NOT PROOF — a battery
+falling 40 → 3 still prints a non-zero count — and one pinned TOTAL rots the moment a sibling grows.
+
+**Handshake — the verdict sets a module-level flag, and the dispatch refuses when it is unset.**
+Set the flag as the self-test's last statement, after its success line prints; the dispatch must
+SAY the self-test never reached its verdict. ⛔ An exit code is not a handshake. Without this a
+`return` above the verdict prints nothing and exits 0, and a perfect floor never runs either —
+the two holes are ORTHOGONAL, so close both.
+
+**Copy a landed one — ⛔ never import one.** `scripts/check-agent-model-declared.mjs` carries all
+three parts (`SELF_TEST_BATTERIES`, `SELF_TEST_BATTERY_FLOOR`, `selfTestReachedVerdict`); every
+self-test must keep running standalone as `node scripts/<x>.mjs --self-test`, so a shared assertion
+module is one point of failure for every instrument at once.
+
+Both non-handshake shapes, and how to classify and probe your own:
+`docs/audits/2026-09-self-test-shape-census.md` and the `scripts/measure-self-test-floor.mjs` docblock.
+
+---
+
 ## Post-Task Checklist
 
 1. `pnpm test` — verify nothing broke. Touched a type-check-covered package? `pnpm typecheck` too.
 
@@ -127,6 +127,25 @@ instrument's header that is **not** a hold: a comparison against a missing retur
 to be false, and nothing detected anything. It belongs on the repair list beside the
 4 DEFEATED, not among the 165 that held.
 
+#### The two non-handshake shapes — the worked examples, held here
+
+`AGENTS.md`'s `--self-test` section points here for these two rather than carrying them itself.
+They are the two shapes an author reaches for when the dispatch is supposed to notice an early
+`return` above the verdict. Neither notices anything; the difference is only in the exit code:
+
+```js
+// the DEFEATED shape — 4 rows above carry it
+process.exit(selfTest());                 // early return → process.exit(undefined) → exit 0
+// the ACCIDENT shape — this row carries it
+process.exit(selfTest() === 0 ? 0 : 1);   // early return → undefined === 0 → false → exit 1,
+                                          // ZERO BYTES printed; the arithmetic did it, not a check
+```
+
+⛔ Neither is a handshake. **A handshake is the dispatch reading a module-level flag the VERDICT
+set, and SAYING so when it is unset** — which is why a non-zero exit alone cannot be scored as a
+hold, and why the ACCIDENT row is classified apart from the 165. The 159 HELD rows that answer the
+mutation do it in one sentence: `selfTest() returned without reaching its verdict`.
+
 ### What the 9 NOT MEASURED are, and what they are not
 
 They are limits of the probe, published per row rather than folded into a verdict — 9
 
@@ -1361,7 +1361,26 @@ export const CEILINGS = new Map([
   // and holds the gate's LEVEL axis unchanged). AGENTS.md is not a CROSS_FILE_MOVES
   // destination, so no `ruledRaises` record applies. Landed count, headroom 0, same
   // convention.
-  ['AGENTS.md', 1075],
+  //
+  // 1075 → 1099 (card #15410): the `--self-test` shape section — nothing in this repo
+  // said what a self-test must look like, so every author re-derived it and the two
+  // holes (no assertion floor; no verdict handshake) kept being re-entered. Landed as
+  // the two judgeable clauses plus one pointer line: the floor pins battery NAMES with
+  // a per-battery minimum (never one total), and the verdict sets a module-level flag
+  // the dispatch refuses on when unset. The narrative, the census numbers and the two
+  // worked non-handshake shapes are NOT here — they were moved into
+  // `docs/audits/2026-09-self-test-shape-census.md`, which the pointer line names
+  // alongside the `scripts/measure-self-test-floor.mjs` docblock. +24 lines, the
+  // measured count of the compressed section including its heading, against a first
+  // draft that measured 41; the section has 0 lossless rewrap headroom at the
+  // surrounding ~100-byte prose width, and the ruling's single pointer line is two
+  // physical lines only because one would be 170B against this file's own 120-byte
+  // per-line budget. Maintainer ruling, verbatim and untranslated:
+  // 「135 同意」 (decision batch #135 item 4, 2026-09-15, presented as 「B,目标 ≤ 24 行,
+  // 并按实测行数抬上限」; recorded by the director on #15410 comment 5682595374, which
+  // also refuses option D). AGENTS.md is not a CROSS_FILE_MOVES destination, so no
+  // `ruledRaises` record applies. Landed count, headroom 0, same convention.
+  ['AGENTS.md', 1099],
   // #9965: root CLAUDE.md is the other repo-root instruction file — same read
   // path (every seat session), same governance (Prime Directive #14). It is
   // structurally growth-prone in the way the ratchet is built for: it exists to