[midhatn]

On Windows, the shutdown observer's two-second connect limit can expire before the operating system returns an actual localhost connection refusal. After the observer has accepted a gateway instance and observed the terminal marker, allow up to five seconds for each blocking operation, clamped to its remaining outer allowance. Ordinary polling keeps its two-second bound. Only an actual connection refusal establishes closure; timeouts, resets, changed identities and stalled responses still fail. Record the selected/effective bounds and underlying error details.

Preserve the original failed physical observations and exact predecessor source. The owned-loopback diagnosis reproduced three two-second connection timeouts versus three five-second actual refusals; deliberately stalled responses remained failures. Ten component tests passed, followed by two affected-test checks after removing a scheduling premise from the test clock. These are not twelve distinct tests, and they do not requalify an earlier physical run.

Also record the completed installed-Hermes SDK active-compiler gate. A real running CTest operation was interrupted, its children were joined, interrupted feedback was committed, admission stayed closed during the foreground hold, and a distinct compiler call completed the same task within its original deadline. Independent closed review passed 61 checks and preserved all 114 original evidence files. The SDK/compiler were real; provider replies, foreground/resource observations and inference-owner actions were scripted.

Update the roadmap and handoff to make actual IQ3_S with physical foreground activity the next bounded gate. Loss while a tool is active, loss of the whole host, successful OS-reboot continuation and full autonomous media qualification remain open. No production runtime module or published 0.4.0a1 asset changes.

Retain PR #40's successful exact-head hosted validation and ordinary protected merge in the canonical fixture record, alongside the earlier failed and incomplete attempts. This follow-up includes that actual main merge in its ancestry.

Validation: source/component reviews and the closed SDK audit described above. Fresh verification of 3a05cac72065562a796acb85acdd19f0c304d7c8 checked all 1,428 raw Git blobs against the checkout and package snapshot; all 77 runtime modules and 88 wheel payloads match the retained package. Document checks covered the exact 15-file change, 339 local links, 53 anchors and the source-bound records. The original hosted checks for this new head remain a separate required gate. The predecessor archives, source pins, scope and citations are retained in the repository.