[StepenkoAnatoli]

Documents-only: the break-test skill (pass 8) run against this repository at 195b2ca, one commit, one new report.

What changed

• docs/build-report-2026-10-11-break-test.md (new) - the pass-8 report: the probe set, the single finding, what was tried and abandoned, and what was not probed.
• docs/README.md - the reading map gains its row.

No kit code, no tests, no corpus.

Why

The break-test skill was run on this repository. Its deliverable is the report; the previous pass's report lives beside it (pass 7, 2026-10-07), which this one continues.

What it found

• Baseline: three full-suite runs at 195b2ca, each 1782 passed, 0 failed, empty failure set, 4 unsupported (the browser-guard/browser-transport LIVE tests; no Chromium on the probe host), exit 1 per ADR-0108. Root gate: preflight PASS, handoff OK, doctor exit 1 with its one by-design blocker.
• The gaps pass 7 left open are closed. Pass 7 could not isolate the network (unshare -n refused). Here unshare -rn works, so the suite ran with no DNS and no outbound sockets: 1782/0/4, identical to baseline. A root-in-userns control (unshare -r) was also identical. Pass 7's other gap - no dedicated repeated-run flakiness programme - is closed by 70 repeated runs of the seven timing-sensitive files (concurrency, http-linear, proxy, large-response, panel, stdout-flush, browser-guard): zero flakes. Seven full passes in total (three baselines, root, no-network, TZ=Pacific/Kiritimati, two simultaneous suites) with the same 1782/0/4.
• One finding, not fixed:

  [F1] handoff from inside a project's research/ folder says nothing has been collected and advises collecting there (Low, Possible)
  Repro: cd <project>/research && node "$HOME/.agents/research-kit/bin/handoff.mjs" prints "nothing has been collected in this project … Collect it here", while the parent directory holds a complete corpus and prints OK. Harm is bounded - the suggested command refuses to spend (research.mjs there exits 2, no plan.json) - but the diagnosis is wrong and points at a metered action. Not fixed on purpose: detecting "this cwd is the project's research/ folder" is new behaviour, and the freeze (ADR-0117) makes it the owner's decision. The report carries a recommended minimal change.

• Everything else probed produced a named diagnostic rather than a crash: no HOME, minimal env -i, no git on PATH, read-only HOME and research/, RESEARCH_KIT_HOME naming a regular file, a taken panel port, a nonexistent --project, a missing capture, a CRLF-smudged corpus, Hebrew + space paths, a 272-character path, unicode TMPDIR, ulimit -n 64, core.autocrlf=true.

Verification

• Verified: selftest.mjs on the committed head in the real checkout - 1782 passed, 0 failed, failure set empty, 4 unsupported, exit 1, 172.0 s, 87 files, Node 22.22.3.
• Verified: the four CI extra steps by hand - release example 6/6 PASS; the three Node<->Python conformance pairs agree per vector; timeline.mjs (60 events) and export-warc.mjs (55 records, 27 captures) regenerate; git status --porcelain empty. preflight PASS (0 blocking, 1 hygiene/brief-unstamped warning), handoff OK (33 ledger entries, 27 evidence rows, chain verifies).
• Verified: the two doc-facing test files - selftest.mjs repo-hygiene support-policy 31 passed / 0 failed - and git diff --check clean.
• Not run: Windows, macOS, Node 24/26, and the browser LIVE tests (no Chromium on the probe host). CI covers the first three on this PR. The browser LIVE tests stay the report's top open risk.
• Environment note: the probe host differs from pass 7 in one useful way (unshare -rn permitted) and is otherwise the same kind of container; the report states both, and its "Not probed" section is explicit about the rest.

What I got wrong

Two probe defects of mine, both recorded in the report and not counted as kit findings: the first network-off run was started without ip link set up on loopback (its FAILs were mine; discarded, re-run clean), and one preflight invocation used a kit-relative path from a project folder that does not contain the kit (Node MODULE_NOT_FOUND).

Merge

The report is documentation only and the branch carries no kit change; merging is the operator's call. No gate override of any kind was used.

[StepenkoAnatoli]

CI record for 9b7eb1c: all seven required jobs passed on the first push (run 38125774529) - platform (ubuntu-latest) 3m10s, platform (ubuntu-26.04) 2m53s, platform (windows-latest) 8m26s, node (24) 3m22s, node (26) 3m27s, archive tree 3m16s, aggregate suite pass. Note for the record: the Windows leg was the 2026-10-11 flake host in the report's risk 1 (the LIVE browser-guard meta-refresh test, which passed here). The four browser LIVE tests remain the report's top open risk; this run did not exercise them again.