test(tui): restore gap replay and saved-answer coverage

[lidge-jun]

Summary

Ordinary dependent draft above #555. Three TUI test files and one historical recorded-frame JSON fixture only; no product/runtime/provider code change.

• Restore gap-preview, replay, saved/compatibility/null-answer and current-owner regression coverage.
• Validate declared HTTP method/path/query/owner tuples before custom replies, then assert complete ordered request counts and an independent violation ledger.
• Exercise saved/absent/unavailable variants for historical print frames, retired assistant/thinking tombstones, separate tool/message suffixes and actual Screen flush followed by an owned saved-answer correction while B remains active.
• Preserve historical backendHead 2a043b14a531ee1db9c2d51b25e1c86f71daac54; clarify that these are historical consumer frames, not a fresh backend/provider capture.

Source 5450d14aa3fcc2492f9e59b643d823be5e654105 cherry-picked unchanged. All four hashes match the independently accepted R2 review, which closed the four earlier test-adequacy findings.

Original TUI verification (05a0b8f36)

184 tests /11 files passed: 116 in the three changed TUI files plus 68 in eight named answer/read/identity/text/geometry/lifecycle/print neighbors. No full local suite.

A separate read-only TypeScript AST check proves all 47 original test declarations remain byte-identical (58 original expanded cases). The 52 restored cases plus six added recorded-fallback variants remain. Historical scenario/identity/frame arrays match the recorded donor, and all reviewed expected annotations/metadata match the accepted source fixture. Isolated homes, bounded output, completed owned-process cleanup and no skipped cases.

These checks use captured renderer output, an ANSI terminal model and xterm geometry. They do not certify an OS PTY, Windows/ConPTY, fresh provider execution, packaged Electron or the separate unready process harness. That harness is explicitly excluded. Hosted CI is tracked on this exact head; existing parent failures and final QA gates remain open.

Dependency

Manual chain: #551 → #552 → #553 → #554 → #555 → this PR. Merge bottom-up only after required per-head review/CI/impact-QA. No native stack, old-donor closure, auto-merge, release or deployment.

Windows physical-fixture repair (9ad655d)

Carries the reviewed session-test repair from 8269b2679 in lower #553 and normal parent merges upward. Real held children replace an autonomous 2-second exit; physical factory cleanup uses the unchanged production default budget while synthetic 100ms fixtures remain unchanged. Exit/active-count assertions run immediately at the production return boundary, before an independent close wait; a foreign PONG sentinel verifies sibling preservation.

Fresh scoped verification at this head: 52 tests /2 files passed. The only change from this PR's previous head is claude-sdk-session.test.ts; production termination policy and the complete 15-file Windows selection are unchanged. Actual new-head Windows and Node CI are required and tracked separately. Earlier test/runtime receipts above are not relabeled as fresh runs on this head. No force-push, rebase, skip or gate waiver.

Rendering and bootstrap repair update (37cd576)

Includes the reviewed Node A repair from 916653397: behavioral badge/final-render proofs, actual Mermaid callback dispatch, settings-failure/location ordering and AST-verified API counts. These repairs are present in lower #551 and propagated normally; the required existing replacement tests are retained there. Each layer freshly passed 60 tests across six named files after that update.

Also includes the reviewed Node B repair from be4f5b3b: observable reconnect/history/sidebar readiness, declared HTTP ledgers and an actually-started stale-A request. Model-free/cache/request-bridge repairs start in #551; the newer redirect fixture is repaired only in its existing #554 layer and descendants. No new test is introduced into an earlier layer merely to apply a patch.

Fresh bootstrap verification at this head: 70 tests /7 files passed. The test-only repairs preserve production code, process policy and Windows selection. Exact-head hosted Node/Windows results remain mandatory; previous Windows successes are not relabeled as new-head CI. No whole-suite local run, force-push, rebase, skip or approval bypass.

Shared disconnect repair (87b46ec)

This head carries the shared lowest-layer repair through normal parent merges: disconnect cleanup uses the existing synchronous UI import, and a same-stack exact-effect regression preserves every prior oracle. Fresh Node22 focused verification at this head: 80 tests across 5 files, all passed with closed processes and removed owned roots. Historical CI failures remain anchored to their original heads; fresh hosted CI, other runtime reviews and applicable QA are still required. This update is not final merge approval.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}