fix(mac): transcript visual restoration — reuse corruption, card chrome, hover meta

[SergeSerb2]

Problem

The AppKit-ported chat log looked broken and cluttered:

1. Row-reuse corruption (the "jank") — clearContent() used removeArrangedSubview, which only unmanages a view; it stays in subviews, drawn at its stale frame. Every reconfigure stranded previous content on screen: stray bars, ghost rows, decision options drawn on top of each other. Separately, recycled NativeWrappingTextFields kept the previous row's preferredMaxLayoutWidth, so reused fields wrapped at ~70px — approvals, checkpoints, and commands rendered one word per line.
2. Chrome loss from the port — tool/approval/checkpoint/plan/reasoning/decision rows lost their card backing entirely (bare text over the scenery photo), the user bubble lost its hairline stroke, the transcript fade veil under scrolled-away tails was deleted, and hover-revealed timestamps/actions became permanent lines under every message.

Fix

Reuse correctness

• Strands removed explicitly (removeFromSuperview on recycle, recursive for stacks).
• Wrap width reset on field configure + deterministic single-line intrinsic before a width is assigned — kills both the narrow-wrap collapse and the stale-height overlap.

Chrome restoration (matching the retired SwiftUI renderer's TranscriptStyle vocabulary)

• Card backing for non-prose rows: radius 12, state-tinted fill (red 8% failed, accent 8% running/decision, quiet 5% settled), 0.5pt hairline, 2.5pt accent rail on decisions and failures.
• User bubble regains its accent-tinted hairline stroke.
• Disclosure headers get the same accent icon tint as plain headers.
• ChatTranscriptFadeEdge (gradient-masked material) restored under scrolled-away tails.
• Content column 800 → 840pt (old renderer's width).

Declutter

• Timestamp + Copy/Edit/Retry/Copy Group fold into one hover-revealed meta line per row; assistant "Responding · time / Copy Response" chrome hover-reveals once settled (stays visible while streaming). Buttons remain enabled — keyboard, VoiceOver, and programmatic activation unaffected; hidden chrome ignores mouse hits via HoverChromeStackView.hitTest so invisible buttons can't swallow clicks.
• Card-backed rows include their insets in the height estimate, so first paint no longer clips card content until the 50ms height commit.

Verification

• Full mac suite: 1107 + 296 + 36 tests, all pass (pnpm run test:mac).
• pnpm run verify --all: 5 steps pass.
• Default UIProbe sweep: passes with only the three activity-dock mount-check failures that already fail on main; before/after captures show cards complete on first paint, full-width wrapping, no ghost content, hover chrome hidden at rest.

Follow-ups (not this PR)

• The three pre-existing activity-dock probe mount-check failures on main.
• Code-block bottom-edge clipping in one mock scene (pre-existing).
• Estimate calibration for multi-line approvals (first-paint pop is reduced, not eliminated).

🤖 Generated with Claude Code

[SergeSerb2]

Follow-ups from the PR description are now fixed in this branch:

• Activity-dock probe failures: the three mount checks (activity-dock-not-mounted, tool-promotion-not-mounted, activity-dock-thinking-not-mounted) failed on every sweep since Gate F because the probeSurface registrations lived on the retired SwiftUI scroll view. Both dock surfaces (pinned + detached) now register from the ChatScreen overlay. Default sweep now passes clean — UIProbe: done with zero failures.
• Code-block measurement: replaced per-line font math (undershoots TextKit's typographic line height) with the laid-out text's actual usage bounds; fragment measure path un-clamped from the stale 800pt. The one remaining "clipped" mock capture turned out to be an honest mid-stream render (message cancelled before the closing fence streamed), not a measure bug.
• Estimate calibration: approval/decision/checkpoint rows now count their caption, choice, and control lines in the height estimate, so first paint no longer clips card content.

Full mac suite green (1107+296+36); the one preferredQuickChatProvider failure seen in an intermediate run is a pre-existing async-seed flake — passes on rerun with no diff change.