fix(desktop): SSH aliases follow port changes#24449

Problem

Adding an SSH environment through a config alias saves the resolved hostname and port as overrides. After changing the alias port, terminal SSH succeeds but the desktop keeps connecting to the old port.

Change

Preserve the entered alias and explicit username/port overrides during registration and reconnect. Discovered-host selection uses the same target parser as manual entry without converting config defaults into explicit overrides. Reconnect no longer rewrites the saved target or invalidates routing consent merely because config resolution differs.

Existing saved profiles retain their values; previously pinned connections need removal and re-addition once. Explicit overrides still take precedence. No profile migration or new persisted fields are introduced.

Scope and approval

Fixes https://github.com/pingdotgg/t3code/issues/7568

GitCafe: #7568

Triaged bug via the automated issue flow. Registration, reconnect, and discovered-host selection are three paths that persist or consume the same SSH alias intent.

Verification

• Regression tests failed before the implementation: registration saved resolved defaults, and reconnect replaced an alias target with a resolved target. Both now pass, including repeated prepares and unchanged routing consent.
• Final focused command: vp test run packages/client-runtime/src/connection/onboarding.test.ts packages/client-runtime/src/connection/resolver.test.ts packages/ssh/src/tunnel.test.ts: 25 passed. Test output.
• Scoped connection/settings tests: 374 passed; client-runtime and web typechecks, targeted lint/format, and git diff --check passed.
• Real Linux Electron under Xvfb: baseline alias reconnect failed against the old port after config/listener change; fixed app connected through 24570 → 24571 → 24572, preserving the connection/environment identity, one route, and the saved alias with a blank port.

Evidence

Linux Electron/Xvfb, Chromium, DPR 1, reduced motion, CSS viewport 1440×900 for the runtime pair below. The final-head runtime recording uses a different isolated local backend port from the baseline; it demonstrates connection behavior, rather than a pixel-matched comparison.

After runtime transcript records the saved target, resolved ports, stable connection/environment IDs and route count. Runtime driver.

Discovered-alias recording, runtime transcript, and driver show blank username/port remaining null while the resolved port changes from 24570 to 24571 and reconnect succeeds.

Proof matrix · Before index · Before recordings, frames, logs and drivers.

Final after index and narration · After matrix · Complete after assets and drivers.

Explicit username-only, port-only and combined overrides connected against deliberately conflicting config defaults before and after another defaults change. Allowlisted runtime transcript · Driver.

Supplemental picker-control comparisons below use the same query and suggestion at each size/theme. These are crops in addition to the whole-surface recordings and frames in the indices. Zero differences are amplified 8×; the second mask marks max RGB-channel delta >16/255. Background routes/provider state and translucency differ, so these do not establish a full-frame visual pass.

What else changed in the fixtures: before failure versus after success, resolved-address subtitles versus saved-alias subtitles, route counts during discovery testing, local backend port in the primary runtime recording, provider availability, retry countdowns, carets and route-added toasts. No layout or CSS changed in the patch. These differences prevent a “nothing else changed” claim.

Unverified: macOS execution; three Before V1 interaction cells; complete blank-and-explicit variants in every V2 cell; strict settled full-frame comparisons and full-frame diff pairs. After has four supplemental V1 and four representative V2 recordings, but only 15/24 static pairs settled identically; nine caret/toast pairs failed, including all V1 result pairs. Supplemental V1 reuses the existing alias through deduplication rather than creating four fresh environments. Runtime proof is affirmative; the full visual matrix is incomplete.

Independent proof audit supports R1/R2/R3 within the disclosed UI/IPC scope and reproduced the 390-dark result pixel-identically, but returned INSUFFICIENT for the complete visual matrix. It also leaves independently verified loaded-build provenance and complete reconnect narration unclosed. Audit inspection assets.

Written with Claude Opus 5.5 through the Claude Agent SDK; triage, review and proof with GPT-6.1-sol through the Codex harness.