test(gui): pin the target banner to the process, not to the field

[donislawdev]

Why

Stacked on #67 (--base is the chunk-1 branch, so review that one first).

test_a_gui_session_keeps_the_target_banner_honest only ever moves the expression: the user types something that matches nothing, then something that does. The case a tester actually hits moves the other end - the field is left alone and the targeted program exits, or a harness restarts it onto a new pid.

Nothing covered that, and the handoff note had concluded from reading the code that the verdict was taken only at session start. It is not: App._refresh_target re-reads targeting.matched on every tick, and only the APPLY half is gated on the expression having changed. Correct behaviour, zero guards - which in this project is how a true sentence turns into a false one a few sessions later.

The guard

Real engine, real resolver, SyntheticDivert, injected FakeTable. It empties the socket table under a running session and asserts _applied_target did not move (otherwise it decays into the old test), then that the banner rises, then that it comes back down when the process reappears under a new pid with the same name - the name-targeting recovery measured 2026-07-28 against a real capture.

Mutants

Three, all caught, each on its intended assertion:

mutant	caught by
_refresh_target reads the verdict only when the expression changed (the assumed behaviour)	a target that died must raise the banner
ProcessTargeting.matched forced to True	a target that died must raise the banner
the banner is never taken back down	the banner outlived the problem

Verification

• python -m pytest tests - 718 passed, 0 failed (elevated shell).
• python smoke_gui.py - OK.
• Test-only, so CHANGELOG-INTERNAL.md only (convention 39).

🤖 Generated with Claude Code