fix(terminal-core): cache-lifetime protocol state (backlog 003) + ED3 resize-wipe auto-repair

[rockyway]

Summary

Two independent bugs, both found on the same live pane (pn-15kpdw3ls) in one debugging session, and both rooted in the same class of problem: state TermFlow observes from the PTY byte stream gets lost or corrupted specifically when a pane spends time as a cached-but-unmounted background tab/pane.

Fix 1 — backlog 003 (Shift+Enter/Ctrl+J permanently broken): ConPTY's one-shot CSI ?9001h (Win32-Input-Mode) and Kitty's keyboard-flag handshakes were only observed while a pane was mounted (CSI parser handlers registered in mount(), torn down in unmount()). A handshake arriving while backgrounded was missed forever. Moved the state-mutating handlers (Kitty >u/<u/=u/>m, Win32 9001 on ?h/?l, DECSTR) to cache-entry lifetime — registered once at Terminal creation, disposed only when the tab is actually closed — mirroring the existing dataDisposable/exitDisposable pattern.

Fix 2 — ED3 resize-wipe auto-repair: codex (ratatui) answers a resize with ESC[2J ESC[3J (erasing scrollback) then re-emits its own retained transcript, capped at ~1000 lines. The backend runs an independent vt100 parser per terminal whose scrollback survives 2J/3J for already-scrolled history (proven by a new regression test) — the data was never actually lost, only the live browser-side view was clipped. Added a new GET /terminals/:id/full-scrollback endpoint exposing that authoritative buffer, detect a CSI 3J arriving shortly after our own background-reactivation convergence resize, and silently repair the live view via term.reset()+write() — the same mechanism hydrate() already uses for reattach.

Full design doc + implementation plan in termflow-fabric:

• docs/superpowers/specs/2026-07-24-protocol-state-and-resize-wipe-fixes-design.md
• docs/superpowers/plans/2026-07-24-protocol-state-and-resize-wipe-fixes.md

Process

Followed brainstorm → design → plan → TDD implement → internal dual-agent review (spec-vs-implementation audit + architecture/lifecycle/security review, run in parallel). The architecture review caught 3 real issues in the ED3 repair path (independently corroborated in part by the spec audit), all fixed in a follow-up commit:

• runEd3Repair could clobber actively-streaming live output with a stale fetch (missing the idle-settle re-check its sibling reconcileSnapshot already has) — fixed with a proper settle-gate + re-arm.
• ed3RepairTimer wasn't cancelled in unmount(), so a scheduled repair from an abandoned engine instance could fire against a since-remounted pane — fixed to match fitTimer's teardown.
• convergenceResizeAt was stamped unconditionally on every activation (narrow false-positive risk on an unrelated clear shortly after switching tabs) — narrowed to stamp only when a real backend resize is actually pending.

Test plan

• packages/terminal-core full suite: 470/470 passing (was 461 baseline)
• Root bun/jest suite: 633/633 passing
• Rust: cargo build, cargo test (191/191), cargo clippy (no NEW warnings — pre-existing baseline findings in unrelated files confirmed unchanged)
• tsc --noEmit: 0 errors
• Production build (bun run build): clean
• Manual smoke (real ConPTY timing, can't be substituted by unit tests): background a tab before codex finishes starting, switch back, confirm Shift+Enter now inserts a newline instead of submitting — flagged for the user to verify live, not run by this PR

🤖 Generated with Claude Code

[rockyway]

External review round (agy + codex, via agent-research)

Both reviewed independently in parallel, focused on TerminalEngine.ts's mount()/unmount() changes, cache.ts's new fields, and the new api_server.rs endpoint.

agy — 1 finding: runEd3Repair's re-arm branch could schedule a new ed3RepairTimer after unmount() already ran (if getFullScrollback was in flight when the pane was backgrounded), leaking a retry loop against an abandoned engine instance. Fixed with a !this.container guard at both checkpoints, matching reconcileSnapshot's existing pattern.

codex — 3 findings:

• High: the repair's reset()+write() never restored the live program's input modes (mouse tracking, bracketed paste, app cursor/keypad, focus reporting) or cursor position — the running TUI still believes those modes are active while xterm silently dropped them. Fixed by appending input_modes_snapshot() (mirroring the existing /snapshot endpoint) and the vt100 crate's own cursor_state_formatted()/attributes_formatted() to the new endpoint's response.
• High: attach() retargeting to a new process mid-fetch wasn't guarded — a stale blob for the old process could commit over a brand-new session. Fixed by capturing attachedProcessId before the fetch and re-validating it's unchanged after.
• Medium: the post-fetch settle check only tested recency, not whether output had changed at all during the fetch window — a slow fetch could let output that arrived mid-request look "settled" purely by elapsed time. Fixed by capturing the exact lastDataAt value before the fetch and requiring it to be unchanged after, in addition to the recency check.

All 4 findings verified as real before fixing (read the actual code/vt100 source, didn't blindly patch), and all fixes verified load-bearing by temporarily reverting each guard and confirming the corresponding new regression test fails before restoring. Full reviews saved to termflow-fabric/docs/review/065-review-pr27-agy.md and 066-review-pr27-codex.md.

Updated test counts: terminal-core 473/473, root suite 633/633, Rust cargo test 191/191 (+1 gated behind integration-tests, compile-verified locally, runs on Linux/macOS CI), tsc --noEmit clean.

🤖 Generated with Claude Code