test(windows): prove one residual cleanup window from supervisor evidence

[BunsDev]

Closes #453

What changes

The native residual regression bounded the time after profile unregistration to 9.5–18 s to catch a window that was reopened or reset part-way through. Wall-clock time cannot do that: a restart at 8.8 s and one slow window both end near 18.8 s, and a slow windows-2025 runner failed #451 at 17.18 s and 18.80 s with nothing reset.

Supervisor (scripts/windows-job-supervisor.cs): the residual window's start is recorded when it opens, after the DeleteProfileW loop returns. A profile cleanup failure now carries OpenCoven.ProfileResidualWindow in its Data: the delete loop's duration, every window start (a reopened window would show two), the moment the window expired, and the window clock's own reading, all as Stopwatch ticks. Production deadlines are unchanged.

Native test (scripts/windows-profile-residual-native.test.ps1): the three full-window cases (persistent, unauthorized, disabled-unregistered) assert on that evidence instead of the ceiling:

• exactly one window opened;
• it ran to its 10 s deadline;
• its clock agrees with its recorded start within 0.5 s, so it was never reset.

The only wall-clock bound left is the existing 30 s total budget. The timing line now prints the delete and window phases, so a slow run shows which phase was slow. A portable check round-trips the evidence through an exception graph and rejects invalid inputs.

The workflow embed is regenerated from the source with renderWindowsSupervisorSource, and docs/phase1-conformance.md carries the new pinned sizes and digests. A repin PR follows the merge, since windows-job-supervisor.cs is governed by harnessAuthority.

Verification

• -CompileOnly and -PortableOnly pass locally (pwsh 7.6.6, macOS).
• Deliberately broken supervisor copies fail the portable check at the right assertion: a describer that accepts no starts → "accepted invalid inputs"; a renamed key → "key changed its contract".
• vitest run src/windows-owned-profile.test.ts src/client-v1-conformance-workflow.test.ts src/windows-supervisor-source.test.ts: 126 passed, 19 skipped.
• Native Windows: this PR carries ci:full for the real residual job. A scratch probe branch (probe/residual-window) also runs the native test against two regressed supervisors on windows-2025: a clock reset at 5 s (lands inside the old 9.5–18 s ceiling) and a reopened second window. Run links follow in a comment.

Coven-Harness: Claude Code / claude-fable-5-1

[BunsDev]

Native Windows evidence

Probe run 38625228667 on probe/residual-window (this branch plus a scratch workflow and scripts/probe-residual-window-mutate.py, neither for merge). One fresh windows-2025 runner per case, the native test invoked from a pwsh step as ci.yml does.

Case	Supervisor	Timing line (persistent)	Result
real	unchanged	delete=1.23s;window=10.07s;windows=1;after-unregistration=10.75s (earlier run)	all 18 cases pass
restart	window clock reset once at 5 s	total=16.75s;delete=1.63s;window=10.06s;windows=1;after-unregistration=16.10s	caught: "window clock disagrees with its recorded start"
reopen	a second window opened after the first expired	total=24.15s;delete=3.97s;window=10.05s;windows=2;after-unregistration=21.73s	caught: "opened its window more than once"

The restart case is the one the issue describes: 16.1 s after unregistration sits inside the old 9.5–18 s ceiling, so the previous test would have passed it. The reopen case ends at 24 s total, inside the 30 s budget, and is caught by the window count rather than by time.

Across the three probe runs the real supervisor measured delete 1.2–1.5 s and window 10.07–10.08 s for all three full-window cases (unauthorized and disabled-unregistered totals 11.6 s). One regressed run showed delete=7.73s on a slow runner, which is the delete-loop variance that pushed #451 to 17–19 s; the timing line now separates it from the window.

The PR's own ci:full run also passed Windows profile residual cleanup and Windows supervisor behavior.

Coven-Harness: Claude Code / claude-fable-5-1