docs: codify recurring Phase 0-3 review findings as agent rules

.claude/agents/c-sharp-dev.md

@@ -135,6 +135,16 @@ Omitting causes `AVLN2100` build errors.
 
 All new public methods need full unit tests exercising all branches wherever possible.
 
+## Pre-report self-review checklist (mandatory before claiming done)
+
+Every phase review so far has flagged the same misses. Run this sweep over the FULL diff (`git diff` + new files) before reporting completion:
+
+1. **Consistency sweep.** Grep the whole diff for every magic literal a new constant replaces — zero stragglers. Every new `await` gets `ConfigureAwait(false)` where the file's pattern uses it. `var` when the type is obvious from the RHS. Blank-line-before-return per style rules.
+2. **Guards on new public APIs.** Every new public method and every `<Name>Factory` method validates reference-type arguments (`GuardAgainst.Null` / `ArgumentNullException.ThrowIfNull`) and degenerate inputs (empty collections) per repo convention. The reviewer has flagged this in every phase — do not ship without it.
+3. **Factories, not ctors.** Records that have a `<Name>Factory` are never constructed via `new` at call sites — including pre-existing records touched by your change.
+4. **Mirror-type parity.** When adding a type that mirrors an existing one (e.g. a new DU alongside `Result<T,E>`), diff its public API surface against the mirror: overload matrix, implicit operators, Task/ValueTask-receiver async overloads. A missing overload here was the only review blocker in four phases.
+5. **DEVIATIONS section.** For refactor/extraction tasks, list EVERY observable difference from the original — concurrency/sequencing, exception types and messages, error paths, delay sites — even ones you consider improvements. Default is to revert to frozen behaviour; a deviation ships only if pinned by a test and declared. Silent "improvements" are defects.
+
 ## Code review checklist
 
 - [ ] Mid-level dev understands in 30 s without comments?

.claude/agents/c-sharp-qa.md

@@ -14,6 +14,7 @@ You are a senior QA engineer specialising in C# 14 / .NET 10 TDD in the AStar.De
 2. **Passing Tests.** All tests must pass. No exceptions. This includes tests not affected by the update.
 3. **One logical concept per test.** A test that asserts more than one distinct behaviour is a design smell — split it.
 4. **Test ADTs, not implementation details.** Test public API and observable behaviour, not private methods or internal state.
+5. **Freeze tests pin the full failure contract.** When pinning behaviour ahead of a refactor, assert exception types and error paths — not only happy-path values. Green happy-path tests are blind to concurrency and exception-type changes. New constants classes get pinning tests in the same commit.
 
 ## Stack and tooling
 

CLAUDE.md

@@ -41,6 +41,12 @@ Read the request literally. "Make the exports work" means fix the bug that cause
 ### Stop immediately when told to stop
 If the user says "DO NOT CHANGE ANYTHING", stop all tool calls immediately. Explain in text only. Do not continue making edits while explaining a mistake — that compounds the error.
 
+## REFACTOR RULES
+
+Behaviour-frozen means NO observable change: sequencing/concurrency (e.g. a sequential `await` loop must not become `Task.WhenAll`), exception types and messages, error paths, timing/delay sites. Green tests do NOT prove behaviour is frozen — tests are blind to concurrency and exception types unless they pin them.
+
+Any deviation from frozen behaviour must be (a) pinned by a test, (b) listed in an explicit DEVIATIONS section of the dev agent's report, and (c) approved before commit. Undeclared deviations found in review are workflow failures, not nits. Default when unsure: revert to the frozen behaviour — never improve silently.
+
 ## Architecture
 
 Targets **net10.0** across all projects. `TreatWarningsAsErrors` enabled in Debug and Release.