feat: v0.22.0 — edition 2024 modernization + layout/render perf + honest lightness docs

[subinium]

v0.22.0 — Rust modernization + optimization release

No new widgets or features. This cycle modernizes the toolchain posture, fixes measured performance issues in the layout/render hot path, and makes every "light" claim in the docs true. Public API surface unchanged.

Toolchain & hygiene

• Edition 2024 for superlighttui and slt-wasm; MSRV 1.81 → 1.88 (let-chains, stabilized 1.88). CI MSRV gate updated.
• indexmap direct dependency removed — it was an MSRV-pin with zero src/ usage. Resolved dep tree: 13 → 10 (no-default-features), 28 → 25 (default) vs 68 for ratatui + crossterm. unicode-segmentation upper bound dropped.
• [workspace.lints] table shared by both crates; lib-only hygiene lints (missing_docs, unreachable_pub, clippy unwrap_used/print_stdout/… family) deliberately stay as lib-target inner attrs so examples/tests aren't flooded. cargo clippy --all-features --all-targets -- -D warnings is clean.
• ~50 nested if-let pyramids flattened to let-chains; core::error::Error impls; OnceLock → LazyLock where const; #[must_use] on remaining builders; 17 never-public terminal methods pub → pub(crate).
• 36 doc_cfg annotations — docs.rs now shows which feature each gated API needs.
• Edition-2024 unsafe env fallout: #![forbid(unsafe_code)] stays for every shipping build; two mutex-serialized #[cfg(test)] env helpers carry audited exemptions.

Fixed

• Startup deadlock on terminals that never answer capability probes — the DA/OSC/DECRPM reply readers gated a blocking stdin read behind crossterm::event::poll() (which reports crossterm's internal queue and consumes bytes from the same descriptor). On hosts that answer nothing (detached tmux panes, script/CI PTYs), apps hung forever on a blank alternate screen before the first frame. Replies are now pumped through a dedicated 1-byte reader thread awaited with recv_timeout — every probe is hard-bounded by its budget (≤180 ms at startup). Verified: detached-tmux e2e renders in <3 s and quits cleanly where it previously deadlocked; input integrity intact (k×3 → exactly 3). Residual on fully silent hosts: ≤1 byte of typeahead, vs. the previous unbounded hang. Adversarially reviewed (concurrency, nudge echo, parser tolerance — all 3 review warnings fixed).

Perf (measured, criterion + v020_perf_audit)

• Intrinsic-size memoization on LayoutNode removes O(nodes × depth) per-frame re-measurement on deep trees; output byte-identical (full proptest suite + 2 new invalidation regression tests); new layout_deep_tree_120x40 bench (depth 14).
• dim_buffer_around rewrite to contiguous row-slice ops: small modal 61.9 → 53.5 µs, large modal 59.2 → 48.2 µs, full scan 106.4 → 55.0 µs (200×60) — fixes the measured regression from perf(layout): replace dim_entire_buffer O(WxH) with modal-aware diff path #228.
• Line-hash double-recompute hypothesis investigated and refuted (line_dirty already makes it near-free; flush/static_200x60 ≈ 114 ns) — honest no-op.
• Guard benches: layout_nested_rows_cols 40.2 → 32.2 µs, full_render_120x40 38.4 → 36.2 µs, animation/churn_200x60 within noise.

Docs

• False "Dependencies: 2" claim in COMPETITIVE_ANALYSIS.md replaced with measured numbers (4 direct required / 25 resolved vs 68). All README taglines (en/ko/ja/es/zh) reframe "light" as dependency footprint + API surface and explicitly do not claim smaller binaries (measured ~1.45× ratatui's hello) or faster cold builds.

Verification (local, all green)

All 11 gates: fmt --check, check --all-features, clippy --all-features --all-targets -- -D warnings, test --all-features (1790+ tests, 0 failed), check --examples, typos, no-default-features, WASM target, cargo hack --each-feature (29 combos), cargo audit, cargo deny check.

Note: the MSRV 1.88 gate is verified by CI (no 1.88 toolchain locally); it is the one gate not reproduced on-box.

🤖 Generated with Claude Code