fix(app): resolve production session compatibility regressions

[henry701]

[ai-agent-generated] This description is LLM-generated.

Scope

Production-only follow-up to #7. Preserve durable steering/queue admission, Stop semantics, rollback replacement, and accurate token accounting. No backend/terminal runtime refactor or upstream feature removal. Audit notes have been removed from the repository and retained outside it.

Fixes

• Provider connection: tolerate a newly mounted integration catalogue that has not hydrated yet. Add delayed-catalogue browser coverage and assert no page errors; connected models remain distinct from available integrations.
• Review tree and filtered file list: reactively bind the viewport when cached panels attach/reparent. Keep this logic in the existing fork helper, with point integrations in the two virtualizers. Regression coverage includes late mounting, reparenting, disconnection and observer cleanup.
• Cached message text: retain indexed highlight spans when normalization creates equivalent reference objects, across native and compatibility rendering. Keep the strict cached-tab first-paint/no-DOM-removal assertions.
• Undo/Redo regression: advance the existing 100ms dialog-close timer before issuing the next shortcut. The portal disappearing does not mean the modal command lock has expired; no command runtime change is necessary.
• History smoke: directly load the session being tested rather than aborting two unrelated documents' initial requests. Retain all history ordering, completeness, no-error, and no-blank assertions. Dedicated cold/cached SPA tab-switch tests remain unchanged.
• Windows CI: replace unavailable Chocolatey ripgrep 15.1.0 with the official release ZIP, pinned SHA-256 verified before extraction.

Additional root causes resolved

• Optional integration discovery must not suspend the entire session when a model picker mounts. Guard the initial resource read; the delayed-catalogue connection form now remains usable before the catalogue response arrives. Model-switch scroll and terminal/composer regressions pass without changing their assertions.
• Context button state changes are atomic so mounting the side panel cannot overwrite the requested Context selection with Review. Cover recorded/live/rollback usage in both legacy and V2 layouts.
• Compatibility normalization now preserves the admitted user payload's agent/model/variant, including before an assistant exists. A later assistant cannot rewrite those recorded choices. This removes contradictory native/compatibility restoration; schema-validated regression and browser manual-selection/submission coverage are included.

Validation (application code unchanged since 0aba3ce6a5)

• Full app browser suite: 249 passed, 3 existing skips, Chromium and Firefox, 5m.
• Full workspace tests: 10/10 tasks passed (8 cached, app/UI revalidated). Earlier uncached affected backend run: core 1,142 tests; opencode 3,627 passed / 22 skipped / 1 todo.
• App: 802 unit tests + 51 browser-condition tests pass. Session UI: 86 tests pass.
• Normal push hook: 30/30 workspace typechecks pass. Separate app e2e typecheck passes.
• Final production tab-switch benchmarks: both pass. Median stable times before/after (ms): legacy cold 162.5/131.3, hot 135.4/119.0; V2 closed cold 171.6/131.5, hot 150.2/119.5; V2 review open cold 170.3/114.5, hot 154.8/113.8. These are local observations, not a performance guarantee; CPU throttling was not changed.
• Official Windows ripgrep archive SHA-256 checked locally; workflow YAML parses. The previous Chocolatey failure was a missing package, not quota. Earlier intermittent Bun patch installation errors have not recurred in the current Windows jobs; no speculative runtime/version upgrade was made.
• Repeated full workspace suite on CI-pinned Bun 1.3.14: 10/10 tasks passed uncached (5m39s). All three HTTP API gates pass on both Bun 1.3.14 and installed Bun 1.4.0: 236 scenarios per mode, 708 total, zero failures/skips/missing routes.
• Hosted Linux and Windows browser suites passed, each with 247 first-attempt passes, 2 retrying cases, and 3 existing skips. The two Linux retrying specs also passed 36 local repeat cases. Retries are disclosed, not represented as flake-free CI.
• Hosted Linux unit/generated-client/HTTP API gates pass. A previous HTTP API run stalled in effect mode and was superseded after more than an hour; the unchanged gate passed on rerun. No quota failure occurred.
• Windows unit diagnosis: the first run hit the 20-minute job-step budget with buffered output. Streamed output then identified three native-process tests hitting Bun’s default 5-second deadline (PowerShell and two Git snapshot tests), with subprocess termination causing follow-on assertions. Those three tests passed after a 30-second budget. Further runs identified the same default-deadline problem in Git repository-cache integration tests, so the final minimal change aligns core’s test command with opencode’s existing 30-second default instead of modifying individual test bodies. All assertions remain unchanged; 1,142 core tests and core typecheck pass locally. The fork Windows job has a 60-minute overall budget; upstream retains 20.
• Windows completed the full core suite with the revised budgets. The subsequent opencode run exposed harness portability issues: path variants must retain their drive when checkout and temp directories are on different drives; immediate EOF includes cold ACP startup; optional terminal capture cannot require an edit when script(1) is absent. The queue smoke now edits through the API when terminal capture is unavailable/disabled, while retaining real terminal edit coverage on Linux. Both paths retain exact FIFO and edited-payload assertions and run locally. 53 affected tests, opencode typecheck, Bash syntax and ShellCheck pass.
• Final Linux gates pass; Windows remains non-blocking per the owner’s explicit instruction. PR merged and local reinstall completed (details below).

Merge gate — final head 05da27e

The owner explicitly authorized merging once Linux browser and terminal criteria pass, without waiting for Windows. Exact-head hosted Linux unit/terminal tests, generated-client verification, all HTTP API gates, browser e2e and typechecks pass. Windows jobs remain non-blocking and are not claimed green. The final CI configuration serializes fork Windows workspace tasks to reduce contention and gives the all-locale ICU initialization test 30 seconds (a prior runner needed 8.4 seconds). No assertions were removed. A prior Windows workspace-sync interruption cascaded into instance teardown failures; three local repeat runs of the workspace/bootstrap suites passed (120 tests). That intermittent Windows behavior remains a follow-up if it recurs.

Result

Merged as 1d4243dd5a803bfc8826e4118e60d27ea8de5e25. Final Linux browser CI: 248 first-attempt passes, one retrying case, three existing skips. All 10 hosted workspace tasks pass, including 3,628 opencode tests; all 708 HTTP API scenarios pass. Installed 0.0.0-production-202609051946 from the merged production checkout with fresh embedded web assets; built and installed SHA-256 match. Backend/proxy services were not restarted. Authenticated live-service smoke testing is left to the owner after restart. Audit notes remain outside the repository.