comment #304
**Golden 34946493637 COMPLETED red — twohost-a/b RCA: CANDIDATE PRODUCT RED (doyle, 09:58Z)**

One paired cell, the web HELPER leg (A `two_host_web_helper_role_a` @ twohost_web.rs:1353 no helper record in 900 s; B `two_host_web_role_b` @ :1035 served 4/5). Every other twohost cell green; floors PASS.

Mechanism (instrumented + source): B's daemon `WEB_SERVE_FOR … outcome=failed reason=DOCS_PORT_DISCOVERY_FAILED: Connection refused (os error 111)`, relayed to A. `webproxy.rs:522-525` (candidate, fold lane f06d4cb6 / 4c6ae0a4, 09-11): the daemon's own `serve_for` opens a local-socket connection to ITS OWN seed socket to ask for the bound docs port and maps every error to `DOCS_PORT_DISCOVERY_FAILED` — no fallback arm (the `may_fall_back` predicate exists only in the CLI's `node_and_port`). The self-connect was refused; the daemon already holds the bound port in-process (`docshost::bound_docs_port`). The last green twohost golden (a9e786b2, 09-10) predates all four #304 lanes and shows no DOCS_PORT string; this run is the path's first cross-host exercise.

Ruling: deterministic candidate defect ⇒ no same-sha rerun can pass; the Windows #309-mechanism red's A/B/C options are moot for S4. Recommended: todlando fix on S4 (+ single-host regression cell + Linux repro) → **S5** → re-gate → new golden (both rig reds already repaired). Referred to the operator via lia. Note: `.spt/preserved/304-handoff/golden-34946493637/twohost-red-rca.md`.