…condary anchors to mcp-permissions

All screenshot tapes used Sleep 1s before and after Screenshot commands
as a timing guard. For the screens affected, Wait+Screen is sufficient:

- help.tape: static CLI output; nothing renders after the text appears
- provider-manager.tape: synchronous list render on navigation
- wizard-screens.tape (frames 1 and 2): synchronous wizard step transitions

The Sleep 1s after Ctrl+Q in wizard-screens.tape is intentionally kept —
that is the alt-screen restore guard (CSI ?1049l), not a screenshot guard.

mcp-permissions.tape gets secondary Wait+Screen anchors instead of Sleep:
- Frame 1: Wait+Screen@5s /3 tools/ confirms tool count has loaded
- Frame 2: Wait+Screen@5s /Server default/ confirms header rows are present
  alongside the tool rows (direct regression guard for netclaw-dev#1424)

Updated stale header comments in provider-manager.tape and wizard-screens.tape
that still described the removed Sleep 1s approach.

…onfig-search baselines

Two separate fixes:

help.tape and provider-manager.tape: restored Sleep 1s before the Screenshot
command. These tapes fire their screenshot right at the start of the tape
(not mid-flow), and VHS may not have flushed all terminal output into its
buffer by the time Wait+Screen matches the first visible line. The Sleep is
a genuine render-settle guard, not timing slop. Removed the post-screenshot
Sleep (that was pure pacing and is not needed).

config-search baselines (approved.png x3): the three config-search screenshot
frames had stale baselines that predated a UI change — they were failing
consistently on every recent Dependabot PR regardless of what code was touched.
Promoted the actual screenshots from CI run 27842198884 (captured with the new
deterministic Wait+Screen anchors from the previous commits) as the new
approved baselines.

…e last capture)

…-byte cmp

VHS's PNG zlib encoder produces different byte streams for visually identical
renders across process invocations (filter-heuristic jitter). cmp -s false-
fails on such runs even though zero pixels differ.

Switch compare_shot_frame to use ImageMagick's 'compare -metric AE', which
counts differing pixels and exits 0 when none differ. This tolerates encoder
jitter while still catching real visual regressions (cursor on wrong row,
missing UI element, etc.). Falls back to cmp -s when ImageMagick is absent.

ImageMagick is already present in CI (used for the existing diff PNG path).

The pixel-comparison fix in run-smoke.sh uses 'compare -metric AE' but
ImageMagick was not present in the Screenshot Regression (Linux) job,
causing it to fall back to cmp -s and false-fail on PNG encoder jitter.

…500)

Set CursorBlink false freezes the terminal cursor but not its on/off state.
The cursor block (~160px) can be visible or hidden non-deterministically
between runs even with blinking disabled. A hard AE=0 requirement false-fails
on this noise.

Allow up to 500 differing pixels — covers the cursor block with margin, while
still catching real regressions: a misplaced cursor row differs by thousands
of pixels, a blank screen by millions.

…-refresh artifact

Root cause: under VHS, Termina emits a full-screen clear (\e[2J) + repaint as
one atomic write when a startup ResizeEvent fires (ConsoleInputSource polls
Console.WindowWidth every 10ms; the PTY size jitters as it initialises). VHS can
sample the screenshot between the clear and the repaint half of that same write,
capturing a blank frame (AE~68000, 1 unique color). The write is atomic from the
app side — real users never see it; only VHS's mid-write PTY sampling does.
Amplified by the Termina 0.12.1->0.14.0-beta.1 bump, which routes spinners
through RenderFrameProvider and raises render cadence.

Deterministic-in-outcome fix (test harness, production code is correct):
- run-smoke.sh: detect a transient blank capture via unique-color count
  (identify %k < 16; blank=1 color, sparsest real frame=7627) and re-run that
  tape up to SHOT_BLANK_RETRIES times. A blank can no longer cause a false
  fail; it was never a false pass (blank != baseline).
- provider-manager.tape: re-assert the list anchor immediately before the
  screenshot (was Sleep-then-capture, which reopened the blank window).

Also widen the cursor-cell tolerance 500->1000 (AE): the shell-prompt cursor
block measures ~493px (one cell); 500 left only 7px margin. 1000 (~2 cells)
clears a single cursor cell while still failing on real text changes.

macOS runners ship bash 3.2 (no associative arrays). The SHOT_TAPE_FRAMES
`declare -A` parsed its [help]=... entries as indexed-array assignments and
aborted under `set -u` with 'line 94: help: unbound variable', failing the
Native Smoke (macOS) job (which loads run-smoke.sh for the light/full modes,
not just screenshots). Replace the associative array with a portable
shot_tape_frames() case function. No behavior change on Linux.

…inism

# Conflicts:
#	tests/smoke/screenshots/config-search-brave-entry.approved.png

The Termina full-refresh artifact also manifests as a PARTIAL frame: VHS
samples mid-repaint so only the top rows land (seen on mcp-permissions-tool-grid
— header + Audience row present, but Server enabled/default and the add/echo/
record-tasks rows missing; AE=19047 vs baseline, yet 4467 colors so the
color-count blank detector misses it).

Generalize the retry trigger from frame_is_blank to frame_needs_retry: blank
(unique-color<16) OR differs from baseline by > SHOT_AE_TOLERANCE. A real
regression also trips the AE branch but reproduces every attempt, so it still
fails at compare time — only latency differs. Tolerance is now a shared global
used by both the retry trigger and the compare gate.

Picks up the upstream fix for Aaronontheweb/termina#311: skip the full-screen
ForceFullRefresh on a no-op resize. This eliminates the blank/partial screenshot
frames at the source (a startup ResizeEvent no longer triggers a [2J clear +
repaint that VHS could sample mid-write). The harness blank/partial retry stays
as defense-in-depth.

…igation fix, termina#314)

…inism