test(reborn): integration-suite restructure — tests/integration/ home, framework disentanglement, single-run coverage

.claude/skills/ironclaw-reborn-testing/SKILL.md

@@ -1,17 +1,17 @@
 ---
 name: ironclaw-reborn-testing
-description: Use when adding or reviewing tests for Reborn behavior — choosing a test tier, covering a bug fix, testing model/tool-choice behavior, touching tests/support/reborn or tests/fixtures/llm_traces, or when a test needs Postgres, Docker, or a live LLM.
+description: Use when adding or reviewing tests for Reborn behavior — choosing a test tier, covering a bug fix, testing model/tool-choice behavior, touching tests/integration or tests/fixtures/llm_traces, or when a test needs Postgres, Docker, or a live LLM.
 ---
 
 # Reborn Testing
 
-Pick the tier first; everything else follows. The repo's tier knowledge lives in `tests/support/reborn/CLAUDE.md` (396 lines — read it before writing harness tests); this skill is the decision layer plus the traps.
+Pick the tier first; everything else follows. The repo's tier knowledge lives in `tests/integration/CLAUDE.md` (read it before writing harness tests); this skill is the decision layer plus the traps.
 
 ## Tier decision tree
 
 1. **Pure logic, no gated side effect** → unit test in the crate (`mod tests` / crate `tests/`).
-2. **A helper gates a side effect (HTTP, DB write, egress body, approval, dispatch)** → you also need a caller-path test driving the real entry point (`*_handler`, facade method, adapter, coordinator). Helper-only is insufficient — `.claude/rules/testing.md` has the bug catalog. Gold standard: `tests/reborn_group_approvals/scenario_gate_then_approve.rs` asserts the approved write **exists on disk**.
-3. **Whole-turn Reborn behavior (submit → runner → loop → reply), deterministic** → the in-process scripted-model harness (`tests/support/reborn/`, run as `cargo test --test reborn_<name>`, zero setup, offline). **Mock only at the vendor-SDK seam** (`TraceLlm`): the real `ironclaw_llm` decorator chain (retry/failover/circuit-breaker) must execute. Mocking at the gateway seam skips it — that's the gateway-seam replay tier's job (`RebornBinaryE2EHarness` / `RebornTraceReplayModelGateway`), not yours by default.
+2. **A helper gates a side effect (HTTP, DB write, egress body, approval, dispatch)** → you also need a caller-path test driving the real entry point (`*_handler`, facade method, adapter, coordinator). Helper-only is insufficient — `.claude/rules/testing.md` has the bug catalog. Gold standard: `tests/integration/group_approvals/scenario_gate_then_approve.rs` asserts the approved write **exists on disk**.
+3. **Whole-turn Reborn behavior (submit → runner → loop → reply), deterministic** → the in-process scripted-model harness (`tests/integration/`, run as `cargo test --test reborn_integration_<name>`, zero setup, offline). **Mock only at the vendor-SDK seam** (`TraceLlm`): the real `ironclaw_llm` decorator chain (retry/failover/circuit-breaker) must execute. Mocking at the gateway seam skips it — that's the gateway-seam replay tier's job (`RebornBinaryE2EHarness` / `RebornTraceReplayModelGateway`), not yours by default.
 4. **Model tool-choice / request-shape is the behavior under test** → recorded QA fixtures (`tests/fixtures/llm_traces/reborn_qa/` + `tests/reborn_qa_recorded_behavior.rs`: ignored live recorder → hermetic contract assertions → hermetic replay). Fixtures must pass `scripts/ci/check-reborn-qa-fixtures.sh` (secret/PII scrub). Never commit unscrubbed traces.
 5. **Browser-visible** → `tests/e2e/` Playwright (`reborn_v2_*` fixtures for WebChat v2). **Live LLM** → `#[ignore]` canary tier; supplemental only, never the PR gate.
 

.claude/skills/ironclaw-reborn-testing/references/exemplar-tests.md

@@ -12,13 +12,13 @@ Living companions to the tier tree in `../SKILL.md`. Each exemplar is a real in-
 
 ## 1. Side-effect proof at the caller
 
-`tests/reborn_group_approvals/scenario_gate_then_approve.rs` — drives a scripted `builtin.write_file` through the **real** stack: first-party runtime → `PermissionMode::Ask` → `TurnStatus::BlockedApproval` → real `ApprovalResolver::approve_dispatch` (lease issued) → `coordinator.resume_turn` → `Completed` — and then **asserts the file exists on disk**. The assertion target is the side effect itself, not a mock's call count. When your change gates a side effect, this is the shape: drive the public entry point, assert the world changed. Re-verify: `ls tests/reborn_group_approvals/`.
+`tests/integration/group_approvals/scenario_gate_then_approve.rs` — drives a scripted `builtin.write_file` through the **real** stack: first-party runtime → `PermissionMode::Ask` → `TurnStatus::BlockedApproval` → real `ApprovalResolver::approve_dispatch` (lease issued) → `coordinator.resume_turn` → `Completed` — and then **asserts the file exists on disk**. The assertion target is the side effect itself, not a mock's call count. When your change gates a side effect, this is the shape: drive the public entry point, assert the world changed. Re-verify: `ls tests/integration/group_approvals/`.
 
 ## 2. The scripted-model harness seam
 
-The in-process harness (`tests/support/reborn/`, spec in its `CLAUDE.md`) fakes exactly one thing: the vendor SDK at the bottom (`TraceLlm`). Everything else — product workflow, coordinator, scheduler, agent loop, the real `ironclaw_llm` retry/failover/circuit-breaker chain — executes for real, and assertions read *persisted state* (filesystem, thread history), never internals.
+The in-process harness (code in `tests/integration/support/`, spec in `tests/integration/CLAUDE.md`) fakes exactly one thing: the vendor SDK at the bottom (`TraceLlm`). Everything else — product workflow, coordinator, scheduler, agent loop, the real `ironclaw_llm` retry/failover/circuit-breaker chain — executes for real, and assertions read *persisted state* (filesystem, thread history), never internals.
 
-- **Right**: mock at the vendor-SDK seam; assert from durable state; `cargo test --test reborn_<name>` runs offline with zero setup.
+- **Right**: mock at the vendor-SDK seam; assert from durable state; `cargo test --test reborn_integration_<name>` runs offline with zero setup.
 - **Wrong**: mocking at the gateway seam (skips the whole `ironclaw_llm` chain — that's the separate binary-replay tier's job); hand-building `TraceStep`s; asserting on internal structs.
 
 ## 3. The declare/enforce contract pair

.claude/skills/reborn-feature/SKILL.md

@@ -87,7 +87,7 @@ For a feature with N endpoints, expect to touch (in dependency order):
 | HTTP | `ironclaw_webui_v2` | route constants + pattern + `*_descriptor()` (use `read_policy`/`mutation_policy`) + add to `webui_v2_routes()`; thin handler over `state.services()`; mount in `router.rs`; **update `tests/webui_v2_descriptors_contract.rs`** (it locks the table) |
 | Wiring | `ironclaw_reborn_composition` + `ironclaw_reborn_cli` | thread inputs through `RebornRuntimeInput`/`RebornRuntime`; attach in `build_webui_services`; pass from `serve.rs` |
 | Frontend | `ironclaw_webui_v2_static` | call endpoints via `apiFetch` in `static/js/pages/*/lib/*-api.js`; consume in hooks. No build step — `node --check <file>.js` to syntax-check |
-| Tests | `tests/support/reborn/` + crate tests | for whole-turn behavior, add a scripted-model harness case (see `tests/support/reborn/CLAUDE.md` — mock only at the vendor-SDK seam); facade changes extend `crates/ironclaw_product_workflow/tests/reborn_services_contract.rs`, and handler changes extend `crates/ironclaw_webui_v2/tests/webui_v2_handlers_contract.rs` |
+| Tests | `tests/integration/` + crate tests | for whole-turn behavior, add a scripted-model harness case (see `tests/integration/CLAUDE.md` — mock only at the vendor-SDK seam); facade changes extend `crates/ironclaw_product_workflow/tests/reborn_services_contract.rs`, and handler changes extend `crates/ironclaw_webui_v2/tests/webui_v2_handlers_contract.rs` |
 
 ## Boundary rules (the guardrails that will reject your PR)