fix(droid-control): preserve text during native typing

[factory-ain3sh]

Description

What

Preserves literal text sent through tctl type on native Linux Wayland. Real Ghostty byte captures reproduce the reported dropped : and Q→Tab substitution on the baseline and pass on this patch. The fix costs one process per character: a warm 52-character prompt measured 403 ms instead of 223 ms.

Why

wtype allocates sequential evdev keycodes for distinct characters. In Reply with exactly: QA ready. Do not call any tools., the first unintended assignment is:

Requested text	wtype keycode	Ghostty physical-key handling	Incoming bytes
:	14	Backspace with printable text	omitted
Q	15	Tab	0x09

Swapping the 14th and 15th distinct characters moves the corruption with their keycodes. A direct raw-stdin probe reproduces it without Droid or the diagnostic wrapper.

How

Routes literal typing and single-character presses through fresh single-character wtype keymaps. -- preserves dash-prefixed text. Chords keep their existing explicit modifier releases; tuistory keeps bulk typing. No clipboard path, custom keyboard encoder, dependency, or pacing delay is added.

Repro Recipe

With Ghostty, cage, and wtype installed, run from the repository root:

TCTL="$PWD/plugins/droid-control/bin/tctl"
S="native-proof-$(date +%s)-$$"
"$TCTL" launch "printf 'READY\n'; cat" -s "$S" --backend ghostty
"$TCTL" -s "$S" wait READY --timeout 15000
"$TCTL" -s "$S" type 'Reply with exactly: QA ready. Do not call any tools.'
"$TCTL" -s "$S" press enter
"$TCTL" -s "$S" snapshot
"$TCTL" -s "$S" close

Expect the exact prompt. The baseline drops the colon and substitutes Tab for Q.

Related Issue

Discovered during CLI-1964 QA. This independently fixes the plugin input driver; it does not change CLI startup or close that work.

Reviewer Guide

Review depth: Standard.

1. plugins/droid-control/bin/tctl: one literal-input owner replaces the bulk call and duplicate character-press branch.
2. tests/test_tctl.py: deterministic process-boundary contract and opt-in real Ghostty bytes.
3. Linux mechanics document and workflow path filter: explain the tradeoff and run the existing test job when tctl changes.

Risk & Impact

• Native typing is slower. Five warm samples per implementation, same isolated Ghostty session and 52-character prompt: baseline median 223 ms, patch median 403 ms. These are local typing measurements, not application latency or cross-machine estimates. Re-measure by timing the recipe's type command against each implementation after warmup.
• Real-emulator coverage is Ghostty 1.3.1-arch2.2 with wtype 0.4-2.2 on Linux, plain mode and Kitty disambiguation flag 1. Other emulators, full Kitty reporting flags, non-UTF-8 locales, and arbitrary custom keybindings were not verified.
• No installed plugin cache, desktop clipboard, authentication fixture, or CLI source was modified.

Verification

Behavior verified @ 7d58d672a6e083a457eed3229a687177b918d3d3:

Control	Result
Bulk → single-character → bulk, same real Ghostty boundary	corrupt → exact → corrupt
New native regression against baseline, plain and Kitty modes	both fail on the original corrupted prompt
Patched suite	all 7 tests pass
Deterministic contract tests against baseline	5 fail; unchanged tuistory routing passes

Coverage includes punctuation/capitals, accented/CJK/emoji/combining text, consecutive calls, Ctrl/Shift chords followed by unmodified text, whitespace, literal dashes, warmup, and stopping on injector failure.

Sentinel test: TctlGhosttyTest.test_native_bytes_across_text_calls_and_chords checks real incoming bytes, not screen reconstruction.

Standard validators: scoped Python unittest suite, Bash syntax, and staged whitespace checks pass. CI runs deterministic boundary tests; native tests require TCTL_NATIVE_E2E=1. All test-created terminal sessions were closed.