Pin managed Herdr installations to 0.9.4

[powerfooI]

Managed Herdr setup now pins 0.9.4 and its five official SHA-256 values. Protocol 22 and endpoint generation 1 remain unchanged. The Windows installer regression checks that the new versioned directory preserves an existing 0.9.0 installation; CLI fixtures and compatibility guidance follow the new baseline. Existing running servers and binaries on PATH are not upgraded automatically.

Refs #411; follows merged #412 and #413. The maintainer requested merging the installer baseline after a clean Codex review and passing checks. The browser and real-agent acceptance checks below remain outstanding in #411. This does not publish a Roamgate release.

Verification

• bun run precommit: format, lint, web build, TypeScript checks and the full suite passed; 3553 passed, 3 opt-in platform/live tests skipped.
• bun run typecheck:bun: passed for all three TypeScript projects.
• Focused installer/bootstrap/CLI suites: 30 passed. Composer and notification suites: 35 passed.
• All five downloaded assets match both the official latest.json SHA-256 table and GitHub release asset digests. Verified the Windows archive includes herdr.exe, ConPTY DLL, x64 and ARM64 OpenConsole helpers, metadata and notices; real Windows extraction and server startup succeeded.

Runtime	Actual execution
macOS 27.0.1 ARM64	Native official macOS aarch64 asset
macOS 27.0.1 ARM64	Official macOS x86_64 asset through Rosetta
Ubuntu 24.04.4 ARM64 / Parallels	Native official Linux aarch64 asset
Ubuntu 24.04.4 ARM64 / Parallels	Official Linux x86_64 asset through Rosetta
Windows 11 ARM64, build 26200 / Parallels	Official Windows x86_64 asset under emulation; PE machine 0x8664

Each row passed an isolated smoke test using Roamgate's actual HerdrClient and EndpointClient: RPC protocol 22, event subscription, generation 1 attach, rendered surface, semantic text/Enter input, pane.read, resize, tab/workspace creation, endpoint reconnect, and passive notification attachment. These are headless transport checks, not browser rendering acceptance.

• The opt-in server/src/connections/windows-ssh-live.test.ts ran against two disposable Windows profiles/sessions: 2 passed, including PowerShell/Unicode input, resize, disconnect/reconnect, and session/generation isolation.
• A real disposable Roamgate backend relayed one custom Herdr notification per connection across Mac, Ubuntu and two Windows profiles, including with a terminal viewer attached. No duplicate relay was observed.
• On native macOS, a 4133-byte API input in canonical mode was rejected with Herdr's 1024-byte-limit diagnostic intact through the Roamgate WebSocket. A sentinel command was neither partially executed nor replayed; subsequent short input succeeded. Existing Composer tests passed draft restoration and error reporting.
• Actual Herdr API state reports labeled Antigravity, Grok, Codex and Pi exercised working-to-idle and repeated-idle/reconciliation paths: one completion per agent through Roamgate's status tracker. This is controlled state-report evidence, not execution of those vendor agents or their screen detectors.

Remaining #411 acceptance checks

• Real browser terminal rendering/input/resize, reconnect and creation, plus visible macOS long-command rejection and retained Composer draft. Browser automation was unavailable in this session: no in-app browser, extension transport failed, and native Chrome/Edge lookup returned cgWindowNotFound.
• Real Antigravity/Grok foreground-prompt transitions with background work still running, and actual Codex/Pi completion transitions; confirm the default Herdr notification source produces no duplicate browser/OS notifications. Controlled API reports do not establish these detector semantics.
• Browser/OS notification delivery, including Web Push with the page closed, as applicable. Only actual backend semantic-notification relays were verified here.

Native x64 hardware was not tested; x64 results above are explicitly emulated. Herdr publishes no native Windows ARM64 asset, and managed setup still rejects that target. No packages, persistent services, firewall rules or permission policies were changed in the guests. Test binaries/configs/sessions were isolated from installed Herdr. Local commands and logs are under /tmp/roamgate-411/; guest test files and temporary servers are removed after verification, with Windows returned to suspended and Ubuntu to stopped.

[powerfooI]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T08:09:42.398055Z	37e8b0c	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Hooray!

Reviewed commit: 37e8b0c96c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".