Add an image tile view to Sliding Puzzle

[mojoro]

Summary

• add a session-only image tile view to Sliding Puzzle, toggled with i, leaving the board, moves, persistence, and reward contract untouched
• pre-encode one opaque native Kitty/iTerm2/Sixel image per board cell — one per tile value plus one per destination for the solved board — once per source/geometry/protocol, so a move reorders already-encoded cells and only the two changed placements are retransmitted
• fall back to Chafa fragments cut from the solved-position preview on terminals without a native protocol, and to numbered tiles while loading, on failure, or when the board area is too small
• generalize the Sixel-only pre-frame wipe to all persistent rasters (pre_frame_persistent_raster_wipe_bytes), covering iTerm2 and non-modal placements, and add stream_modal / history_modal to the raster overlay lists so they match foreground_terminal_overlay_open
• give the board rect one definition (App::content_area, arcade::ui::SHOW_GAME_BOTTOM_BAR) shared by the draw path, mouse hit-testing, and the pre-frame wipe, replacing the hand-rolled layout in arcade/input.rs

Review

• reviewed post-rebase by two independent agents scoped to the shared terminal-image infra and to the game code; findings triaged below
• fixed during review: a clippy::needless_lifetimes error that would have failed CI; an ImageTiles::release leak where a request still in flight when the board closed left its whole encoded set buffered in the channel for the rest of the session; two wipe tests that passed through the modal-close trigger rather than the trigger they named; a duplicated MAX_DECODED_IMAGE_PIXELS; stale Sixel-only doc comments after the rename; tick.rs re-deciding the bottom-bar literal instead of reading SHOW_GAME_BOTTOM_BAR; and two defensive guards (zero tile geometry, saturating_sub in the label centering)
• the repaired tests were mutation-checked: deleting the non_modal_image_changed / protocol_changed terms now fails all three wipe tests, where two previously stayed green

Known gaps, flagged rather than fixed

• Partial re-emission assumes the raster covers its cell rect exactly. The in-place fast path re-emits only the placements that changed. Sixel payloads are sized from a hardcoded 8×16 cell guess and iTerm2 letterboxes via preserveAspectRatio=1, so on a terminal whose cell metrics differ, a tile's raster can bleed into a neighbour that the fast path never repaints. Previously any change re-emitted every placement, which masked it; this is also the first multi-placement non-modal scene in the app. Happy to gate the fast path to Kitty, or re-emit the full set for iTerm2/Sixel, if you'd prefer.
• Kitty image ids are a 24-bit fold of the placement UUID. With 25 cells on screen at once, two can collide (~2e-5 per board, deterministic per build) and one tile would go blank. opaque_replacement_has_same_coverage guards the old-vs-new pair at each index but does not check for collisions between two placements in the same frame.
• No aggregate cap on an encoded set. Hard is 2 * 5² = 50 images, each with a PNG and, on Sixel, a payload bounded only by the per-image SIXEL_MAX_BYTES. Realistic photos land far below the structural ceiling, but nothing checks the total.
• The shipped art is hotlinked from a third party. PLACEHOLDER_IMAGE_URLS hardcodes three signed fastly.picsum.photos URLs. Lorem Picsum has no SLA, no published terms, and no rate-limit policy, and the hmac params come from their signing key — a rotation on their side 403s all three at once and everyone drops to numbered tiles. Every other Arcade asset is embedded at build time with include_bytes!; doing the same here would delete the runtime dependency, the per-session download, and the file:// path together, but it needs an artwork/licensing call from you first. Recorded in late-ssh/src/app/arcade/CONTEXT.md under known gaps.
• artwork/sliding-puzzle/ ships only a README.md and a .gitignore — it is a Compose-bind-mount dev affordance, not a runtime path; no Dockerfile stage copies it.

Verification

• cargo fmt -p late-cli -p late-core -p late-ssh -p late-web -p late-webview -- --check — clean
• cargo clippy -p late-ssh -p late-core --all-targets --features otel -- -D warnings — clean
• cargo nextest run -p late-ssh -p late-core --all-targets — 3395 passed, 8 skipped
• no test touches the network: the file:// tests use temporary directories, and every poll test runs outside a Tokio runtime so nothing spawns

Environment note

make check could not complete on this host: the workspace clippy build needs atk/gdk-3.0 for late-webview, and only GTK4 is installed here (same gap as #551). The scoped runs above cover every package this branch touches — late-webview is untouched — and CI installs libwebkit2gtk-4.1-dev, so the workspace build is covered there. No system package was installed for this PR.