docs: mandate plan-first, red-green TDD, and pre-commit UI validation

Make the three workflow expectations explicit, non-negotiable defaults:
always plan before execution, always use red-green TDD, and always
attempt manual UI validation before committing (skippable only when the
tooling is unavailable, and only with an explicit callout + handoff).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

AGENTS.md

@@ -79,10 +79,22 @@ docs/SWIFT_GUIDELINES.md    Swift coding/testing/patterns/security standards
 
 ## Workflow expectations (user preference)
 
-- **Red-green TDD**: update `docs/RULES_FEATURE_SPEC.md` first for behavior
-  changes, write the failing test, run it (compile failure counts as red),
-  implement, re-run focused tests, then the full suite. Run tests often and
-  fail fast.
+These three are non-negotiable defaults — follow them on every task, not only
+when reminded:
+
+- **Always plan before execution.** Think through and lay out the approach (a
+  written plan / plan mode for anything non-trivial) and confirm scope before
+  editing code. Do not start changing files until the plan is clear.
+- **Always use red-green TDD.** Update `docs/RULES_FEATURE_SPEC.md` first for
+  behavior changes, write the failing test, run it (compile failure counts as
+  red), implement, re-run focused tests, then the full suite. Run tests often
+  and fail fast.
+- **Always attempt to validate the UI manually before committing.** Build and
+  run the app (simulator/device) and visually confirm the change behaves as
+  intended. This step **may be skipped only when such tooling is unavailable**
+  (e.g. the Xcode MCP / a simulator is not reachable in the session) — in that
+  case, say so explicitly and hand the verification back to the user rather
+  than silently skipping it.
 - Conventional commits (`feat:`, `fix:`, `refactor:` …). **Agent attribution is
   required**: every commit an agent authors or co-authors must end with a
   `Co-Authored-By:` trailer naming the specific agent/model that did the work,