session 2026-08-08 (late): harvest proposal — degraded-state shape census (ladder candidate, ~10% retrieval declared)

claude/memory/skill-harvest-register.md

@@ -316,3 +316,35 @@ is evidence about the home, not about the lesson.
 **Status: PROPOSED**, to be ruled with #192.
 
 **No FIX-lane changes were applied this session.** All skill/tooling edits this session were either governed records (`PENDING.md`, the register, memory files) or authorized builds under REVIEWED-100 — none were self-tending edits to a skill.
+
+### 2026-08-08 (late) — ladder candidate: the degraded-state shape census
+
+**Kind:** verification-ladder entry (not a skill). **Firing moment, declared:** on a condition
+the executor must first notice — *"is this suite honest when its subject is missing?"* Per the
+routing table's fourth row this earns **~10% retrieval as a bare ladder entry, and that estimate
+is recorded on the proposal rather than wished away.** The mechanical detector already exists for
+half of it (`run-fleet` now renders the third verdict), which is why only the *method* is proposed
+here.
+
+**The instrument.** To find out whether a check is honest, **drive real degraded states and observe
+the output SHAPE** — crash / named-failure / cannot-assess / pass — rather than grepping the source
+for risky constructs. Earned twice on 2026-08-08:
+
+- The grep census counted **comments and docstrings** (3 of `test_navigate`'s 4 `next(` hits were
+  the executor's own prose *about* the fix); AST corrected 4 → 1. **A construct that CAN raise is
+  not a suite that DOES.**
+- Driving five degraded states — manifest empty · source file missing · db empty · db absent ·
+  manifest unparseable — found **8 crash sites across 4 suites**, and **closing the first six
+  revealed two more in suites the census had already cleared.**
+
+**Companion, same family:** for the discrimination gate's *"a real known-bad where one exists"*
+standard, **git history often holds one**. `git show HEAD:REVIEWED.md` supplied a genuine prior
+register state that lacked three rulings, so the new built-vs-ruled check was proven on a real
+artifact rather than only on a fixture.
+
+**⚠ Against filing it:** the census method is one instance of *"census by mechanism, not proxy"*,
+which is already banked in MEMORY.md's standing preferences. This may be a **restatement**, and the
+honest proposal is that the steward decide whether the degraded-state form is distinct enough to
+earn its own entry or should be folded into the existing one as an example.
+
+**Awaiting:** steward authorization.