Adapt selected Synara and T3 reliability updates

[yaacovcorcos]

Summary

• adapt Synara's bounded Codex child-route, latest-project new-thread, and fixed provider-menu improvements to Scient's existing seams
• add T3-informed terminal selection Copy, a 30-second OpenCode startup allowance, Windows-safe Claude executable resolution, and visible Codex web-search details
• harden OpenCode and Kilo session resume, confirmed-404 recovery, changed-directory fork verification, permission restoration, and concurrent-start ownership
• advance the repo-local Synara review checkpoint without importing the rejected gateway or workspace-refactor architecture

Why

The 2026-07-21 selective donor review found a small set of useful reliability and UX changes in Synara and T3 Code. The implementations here preserve Scient's provider lifecycle, project targeting, terminal behavior, state ownership, and release boundaries rather than copying donor architecture wholesale.

The initial candidate exposed three regression risks: malformed Codex web-search fields could reach string operations, terminal context-menu Copy retained xterm line padding that keyboard Copy removed, and missing OpenCode/Kilo source-directory metadata could falsely claim a cwd move. This PR fixes all three and adds direct adversarial coverage.

The reviewed Synara markdown-heading adaptation is intentionally absent. Comparative browser measurements showed that it worsened the one-second narrow-resize transcript gap from 400 px to 418 px, so it was rejected instead of shipped.

Regression safeguards

• normalize unknown Codex search-detail fields before trimming them
• share terminal clipboard normalization between keyboard and context-menu Copy while preserving indentation, internal whitespace, and CRLF
• reuse provider sessions only after validation; recover fresh only from structured 404s and propagate transient failures
• verify the provider-returned cwd before accepting a fork, preserve the source conversation, and remove only a distinct unusable fork
• prove concurrent durable resumes do not let a race loser abort shared provider state
• cover the shared OpenCode adapter through exact Kilo lifecycle tests

Verification

Validated locally at exact head 8196524ed6585d2d6f8d585327cfb6797830dda2, rebased onto origin/main 71da4e46db150a02192f7e1cb1c995f65281cc0e:

• bun run fmt:check: passed, 1,865 files
• bun run lint: passed, 0 errors and 211 existing warnings
• bun run typecheck: passed, 9/9 tasks
• bun run brand:check: passed
• bun run test: passed, 12/12 tasks; server 2,092 passed and 7 skipped; web 2,868 passed
• stable browser suite: 197 passed and 11 intentional skips
• browser geometry suite: 11 passed
• bun run build:desktop: passed, 5/5 tasks
• bun run test:desktop-smoke: passed, 6/6 tasks
• bun run release:smoke: passed, including native node-pty spawn
• bun run scient:upstream-check --review-check --no-fetch: passed with zero unreviewed Synara commits
• isolated OpenCode 1.18.4 and Kilo 7.4.11 compatibility probes passed without sending model prompts or consuming provider credits
• an exact-head arm64 DMG passed image verification and deep strict code-sign verification

Validation boundaries

• hosted CI is pending on this PR
• Windows behavior has injected platform/PATHEXT coverage but no physical Windows-host run
• the local DMG is ad-hoc signed, has no Team ID or notarization, and is audit evidence rather than a release candidate
• no artifact was installed, promoted, or published

[yaacovcorcos]

Putting this PR on hold. Keep it as a draft and do not merge or resume the broader repair work. Resume only after explicitly agreeing the scope boundary between the narrow donor adaptations/direct regression safeguards and the separate shared provider-lifecycle redesign. Before any push, reconcile the unpublished work and rerun full validation on the narrowed PR.

[yaacovcorcos]

Closing this held mixed intake in favor of draft #107, which reimplements only the three approved bounded improvements on current main: fixed-width provider/model picker, hardened terminal selection Copy, and visible Codex web-search details. OpenCode/Kilo turn completion and all broader provider-lifecycle changes remain excluded for a separate reviewed follow-up. The PR #61 branch and local worktree are preserved unchanged for provenance.