Ctrl+Delete word-forward-delete parity with Ctrl+Backspace

[rockyway]

Summary

• Adds Ctrl+Delete word-forward-delete, matching Ctrl+Backspace across every shell context (PowerShell, cmd, Git Bash/WSL, and Kitty-protocol TUIs).
• Fixes a pre-existing gap: Ctrl+Backspace did not reliably word-delete in POSIX/readline shells (Git Bash, WSL) due to a documented ConPTY limitation where Ctrl+Backspace and literal Ctrl+H both encode to the same byte (0x08).
• Kitty protocol path was already correct/tested; Win32-Input-Mode path was already correct for Ctrl+Delete but untested (now locked in with a regression test); the new piece is an explicit shim in TerminalEngine.ts that sends readline's native word-kill sequences (ESC DEL / ESC d) directly to POSIX/readline shells, gated by a new shellType engine option threaded from the app.

Design / plan docs

• termflow-fabric/docs/design/005-word-delete-shim-design.md
• termflow-fabric/docs/plan/006-word-delete-shim-implementation.md

Test plan

• bun run test:workspace (terminal-core unit tests, 436 passing, incl. new wordDeleteShim.test.ts and the Win32-Input-Mode Ctrl+Delete regression test)
• bun run test (main app suite, 592 passing)
• tsc --noEmit (clean)
• bun run build (terminal-core + renderer, succeeds)
• Manual: Ctrl+Backspace/Ctrl+Delete in Git Bash, WSL, PowerShell, cmd, and a Kitty-protocol TUI (see plan doc Task 5 Step 3) — not yet performed, no GUI-automation available in this session

[rockyway]

Automated review round

Dispatched two independent external reviewers (agy/Gemini 3.6 Flash, codex/gpt-5.6-sol) against this PR.

agy: full approval, no blocking findings.

codex: found 2 P1s + 1 P2, all confirmed on independent verification against pty_manager.rs/StateManager.ts and fixed in 07bf400:

1. isPosixShell default direction was unsafe. The original heuristic defaulted anything that wasn't literally cmd/powershell/pwsh to POSIX. 'default' is a real, persisted shellType value (StateManager.resetToDefaultLayout, and TerminalPane's fallback chain before shell profiles finish loading) that frequently resolves to an actual PowerShell session — PowerShell 7 ships is_default: true and is pushed first in compute_available_shells(). The old default direction would have sent raw ESC-byte garbage into a real PSReadLine session. Flipped to a whitelist of known POSIX profile ids (bash, zsh, fish, sh, cygwin, git-bash, wsl(-*)), defaulting everything ambiguous to not shimmed (i.e. today's pre-PR behavior, not a new regression).
2. The keyup paired with a shimmed keydown wasn't swallowed. Win32-Input-Mode stays active for the whole Windows session regardless of shell, so the unswallowed keyup fell through to that encoder and leaked a stray release record. Fixed via the same uiClaimedKeydownKeys claim the scroll-key shim already uses for this exact problem.
3. shellType was captured once at mount, inconsistent with the established live-getter pattern (smartCopy, enhancedKeyboard) used for exactly this class of "can change after mount, engine doesn't remount" problem. Converted to a getter.

New regression tests lock in all three: wordDeleteShim.test.ts ('default' → not POSIX) and a new engine.wordDeleteShim.test.ts (keydown/keyup pairing across POSIX vs. PowerShell vs. 'default').

Full suite: 446 terminal-core tests, 592 app tests, clean typecheck, all green post-fix.