KKnD closed target helper: controlled fix, new source-observer boundary

Run: scratch/runs/20261106-kknd-closed-target/. The failing predecessor 20261106-kknd-menu-consumer/menu-observer.cjs is preserved unchanged (hash in preserved-predecessor.json). See its consumer handoff for the read-only original guest chain and pre-attachment failure.

The fresh helper skips only recognized closed CDP sessions during readiness or owner evaluation, recording rejection stage, name, message and stack. Live evaluation defects remain fatal. It requires exactly one live slot 0/thread 1, the parent owning endpoint, 512 MB memory and actual instantiated module hash; identity is checked again in the same evaluation that attaches diagnostic counters. The fresh browser adapter preserves original readiness Error objects. No guest code/state writes, production emulator changes, retries or restarts. The new helper source is retained in prelaunch-new-helpers.patch and final-new-helpers.patch; no predecessor files were edited.

All three ordered FINAL releases and independent cleanup passed before acquisition at 12:15:09.609753 UTC (deadline 13:00:09.609753, including cleanup). Before any browser, all 15 exact exported helper controls passed: closed before ready, on evaluate, before/after the live owner; zero/duplicate owners; real VM ReferenceError and non-lifecycle rejection; endpoint/module/attachment-identity negatives. The preserved predecessor still fails its live-owner/closed-sibling control. The passive consumer control and exact browser adaptation compilation passed. Locked full standard/compat build passed. Tested Git SHA: 0ae13a764ecbfb8e81b38ec6033301467d3f0e50, tracked source diff empty, with owned new-file patches retained. All 325 source/module pins and 16 original media files match after execution. Both pre-navigation HTTP pins and all 114 successfully read browser source responses match.

First new failure: the sole browser's retained source-response observer calls response.buffer(), which rejected with ProtocolError: Could not load response body for this request. This might happen if the request is a preflight request. Its catch treats every body-read rejection as actual browser source identity failure and closes the browser. It omitted URL/method/status/redirect information, so the specific response and cause are unknown; the error text does not prove a preflight. No recorded hash mismatch or emulator defect is established.

The actual served standard module and Worker CDP bytecode match 220beaba026d0aabaf25b4193a345ca6d704be64b0812a72d68adb0700ad80f2. The closing identity sample saw one instantiated slot 0/endpoint 1/thread 1, but sampled before the parent running-app binding existed and its strict pass is false. This is not qualified owning-module identity. Later terminal state shows the running original app; personally inspected full/game captures show its splash. No input, menu observation attachment, Return click, consumer evidence, selection, commanded displacement or second gameplay response occurred.

Next boundary: in a fresh helper, record response URL, method, status, redirect chain and body-read error before collection; control bodyless/redirect responses, required-source body failures and real hash mismatches without weakening byte identity checks. The current evidence cannot select a safe exception policy. Then a separately authorized original session can use the controlled menu helper. Do not replay routing or infer a guest CPU/API repair from this observation error.

Browser closed at 12:16:12.698 UTC; independent cleanup passed at 12:16:59: 14 absent owned PIDs, two refused ports, one absent profile, four free locks, 4.63 GB free, within the window. Holder SIGTERM/143 was intentional cleanup. No boxes assigned (ID/state/expiry N/A). Evidence includes failure-boundary.json, closed-target-control.json, helper-checks.json, terminal-executor-identity.json, source-postcheck.json, capture-review.json and final-cleanup.json.