Record green live lifecycle fix campaign and preserved failures

Co-Authored-By: Paperclip <noreply@paperclip.ing>

tests/lifecycle-baseline/LIVE-BASELINE-2026-09-21.md

@@ -2,7 +2,8 @@
 
 Measured on GitHub Actions with real LLM calls on September 21 (America/Chicago;
 September 22 UTC). This supplements the immutable [scripted baseline](BASELINE-2026-09-21.md).
-No production lifecycle behavior was changed.
+No production lifecycle behavior was changed for the initial measurement.
+Subsequent changes and fresh runs are recorded in the [fix report](LIVE-FIXES-2026-09-21.md).
 
 ## Initial measurements
 
@@ -148,4 +149,4 @@ a turn. Earlier cancellation rejects the late session; later cancellation can
 reach the published handle. Deterministic tests inject Stop in that exact gap.
 The existing execution-owned cleanup closes rejected sessions.
 
-Fresh live verification is recorded separately when complete.
+Fresh live verification is recorded in the [fix report](LIVE-FIXES-2026-09-21.md).

tests/lifecycle-baseline/LIVE-FIXES-2026-09-21.md

@@ -0,0 +1,72 @@
+# Lifecycle live fixes — 2026-09-21
+
+This follow-up changes production startup cancellation and repairs test fixtures.
+The [original baseline](LIVE-BASELINE-2026-09-21.md) and its scores remain intact.
+The [fix campaign inventory](baselines/2026-09-21-live-fixes.json) retains each new
+cell outcome, source SHA, definition hash, artifact SHA-256 and cost coverage.
+
+## Changes
+
+- **Native Stop during startup:** Stop could commit after the coordinator claim
+  but before the provider session handle existed. Cancellation was acknowledged
+  without reaching the later session, which submitted work anyway and held up
+  `/new`. The executor now publishes the handle, reads the durable run state and
+  rejects an already stopped session before provider submission. Later Stop can
+  reach the published handle. Execution-owned cleanup and ownership gates remain.
+  Two deterministic tests inject cancellation in this precise gap.
+- **Legacy blockers:** the original prompt omitted the API's required structured
+  blocker and received 422. The model then named itself as unblock owner, which
+  legitimately woke it again. The corrected fixture seeds an unassigned backlog
+  prerequisite and requires a dependency link. The grader reads `blockedBy` from
+  the public GET response and verifies the exact prerequisite remains unfinished.
+  Exact single-run and single-response assertions remain.
+- **Blocked status UI:** the browser assertion uses the fixture's expected state
+  and accepts the Blocked attention suffix. Persisted status is still exact.
+- **Approval proposal identity:** a run saved `welcome-note-approach`, bound its
+  exact revision for confirmation and waited correctly. The generic grader's
+  English key convention misclassified this as a premature final deliverable.
+  The lifecycle fixture now explicitly requires the proposal's reserved `plan`
+  key and its current-revision approval, with no other output before approval.
+  We retain the negative control that an arbitrary final deliverable does not
+  become a plan merely because an agent targets it for confirmation. New controls
+  reject proposal confirmations for the wrong issue, key or revision.
+
+## Fresh campaigns
+
+[35679927390](https://github.com/paperclipai/paperclip/actions/runs/35679927390),
+App `59f01883de1c9447eb8ac68f24ed24cad573fc00`: **37 passed, 3 failed, 40 executed**.
+Both native blockers and native Stop/new/resume passed. The remaining failures
+were two incorrect GET-field assertions and the proposal-identity mismatch above.
+All cleanup passed; no retries. These original campaign scores are not regraded.
+
+[35680906634](https://github.com/paperclipai/paperclip/actions/runs/35680906634),
+App `331e89bb3e40064bb8f51010f176db367588bbef`: **40 passed, 0 failed, 40 executed**.
+All 20 legacy and all 20 native journeys passed, including both narrative variants
+of blockers, approvals and plan revisions, and Stop/new/resume on both runtimes.
+All evidence validated and all cleanup passed, with **zero retries or incomplete
+cells**. The fixture definition changed from the preceding campaign; this is a
+fresh provider measurement, not a regrade or a claim that the original failures
+never occurred.
+
+The normalized report's `complete` flags remain false: those fields qualify
+coverage against the publisher's known catalog/definition hashes and the entire
+multi-suite roster. This branch-only, explicitly selected lifecycle suite does
+not qualify that broader catalog. Its 40 selected cells all executed and passed;
+this is not a claim that every Paperclip E2E suite was run.
+
+## Local validation
+
+- 418 tests passed: native session executor (385) and agent conversations (33).
+- 440 Product E2E support tests passed, including the three new negative controls.
+- Product E2E TypeScript and server typecheck passed. The latter also completed
+  the runner TypeScript/Rust build and generated contract checks. The first local
+  typecheck hit missing `three` dependencies; installation was repaired using
+  repository pnpm 9.15.4, and the tracked lockfile was preserved.
+- No repo-wide test/build qualification is claimed. The separate scripted
+  narrative-authority baseline still contains its recorded failures; this is
+  not the broader removal of prose from lifecycle authority.
+
+Each paid cell is one attempt, using Codex `gpt-5.6-sol`, Chromium, a real isolated
+server/database and either legacy or native execution. A successful campaign is
+not a reliability estimate. Product model-cost coverage is partial/unpriced;
+reported zero is not proof of free execution. GitHub compute is separate.

tests/lifecycle-baseline/README.md

@@ -1,14 +1,15 @@
 # Lifecycle behavior baseline
 
 This suite establishes a measurement before removing narrative/regex authority
-from lifecycle decisions. It changes no production policy. Assertions describe
-intended behavior; an observed failure is retained, not blessed as an expected
-outcome and not fixed by this change.
+from lifecycle decisions. The original baseline changes no production policy. Assertions describe
+intended behavior; observed failures are retained rather than blessed as expected
+outcomes. Subsequent fixes and fresh measurements are recorded separately.
 
 The [initial recorded baseline](BASELINE-2026-09-21.md) preserves passes,
 failures, representative observations and unmeasured coverage. The separate
 [live Actions baseline](LIVE-BASELINE-2026-09-21.md) records actual LLM/browser runs,
-assertion corrections and retained behavior failures.
+assertion corrections and retained behavior failures. The [fix report](LIVE-FIXES-2026-09-21.md)
+records the later cancellation fix, corrected fixtures and fresh campaigns.
 
 ## Run and inspect