Remove HermitCrab's analysis-cascade memo

docs/review/devils-advocate.md

@@ -17,8 +17,8 @@ Challenge the most consequential claim first, one objection at a time, each with
 a `path:line` and a concrete scenario. In this repository the claims that have
 failed before are:
 
-- a HermitCrab performance win asserted without a measured artifact, or a memo
-  key that omits a field a rule reads;
+- a HermitCrab speedup asserted without a measured artifact, or a merge that ignores
+  state read by an analysis rule;
 - a USFM or reference change whose test proves the happy path only;
 - a parity claim about `machine.py` with no checked comparison.
 

docs/review/hermitcrab.md

@@ -1,20 +1,15 @@
 # HermitCrab Review
 
-*Review HermitCrab morphology changes for analysis equivalence, memoization-key
-completeness, retained-memory bounds, parallelism, and hot-path cost.*
+*Review HermitCrab morphology changes for analysis equivalence, state-sensitive
+merging, parallelism, and hot-path cost.*
 
 Governs `src/SIL.Machine.Morphology.HermitCrab/**/*.cs`.
 
-- Treat analysis output as the primary contract. A faster parse, more memo hits, or a
-  successful build does not prove equivalent analyses.
-- When changing analysis-side rules or state, re-audit every field the rule reads
-  against `AnalysisStateKey`. Check freezing, cached hashes, mutable dictionaries,
-  equality, rule counts, non-head counts, feature structures, and stratum identity.
-- Memoized results must represent fully expanded subtrees. Check replay prefixes,
-  deduplication, empty/nogood entries, in-flight recursion, and the separation between
-  sequential and parallel scopes.
-- Do not weaken an existing memo or retained-word bound without measured evidence
-  and a test. Read the current limits from the code.
+- Treat analysis output as the primary contract. A faster parse or a successful build
+  does not prove equivalent analyses.
+- When changing analysis-side rules or state, check every field a rule reads and
+  whether equivalent analyses can be merged safely. Check rule counts, non-head
+  counts, feature structures, and stratum identity.
 - Inspect allocations and retained object lifetimes only in changed inner loops. If the
   change claims a performance improvement, require a reproducible benchmark or measured
   artifact in addition to semantic regression tests.

docs/review/machine-tests.md

@@ -15,8 +15,8 @@ Governs `tests/**/*.cs`.
   when the contract is ordinal identity.
 - For async code, assert cancellation and completion behavior where the change promises
   it; do not hide unobserved tasks.
-- For HermitCrab changes, compare analysis semantics, not only memo-hit counts or
-  execution success. Exercise key completeness, replay, resource caps, and
-  parallel/sequential equivalence when touched.
+- For HermitCrab changes, compare analysis semantics, not only execution success.
+  Exercise state-sensitive merging, resource caps, and parallel/sequential
+  equivalence when touched.
 - Name the test that proves the change. "Where is the test?" is the single most
   common review question in this repository; answer it before it is asked.