perf: trim hot-path allocations in indexing and TUI render

[amittamari]

Concrete, std-only allocation cleanups in the indexing parse loop and the TUI render path. No new dependencies.

Stacked on top of #76 (base = oxc-inspired-phase1). Rebase onto master once #76 merges.

Changes

• Adapters: reuse one decode buffer per file instead of allocating a fresh Vec<u8> for every JSONL line (clear + extend_from_slice). Applies to the Claude, Codex, and Cursor parse loops.
• Results list: compute each visible cell's (text, style) once per frame (compute_cells), then feed the same grid to both the width solver (layout_from_cells) and the row builder (session_row). Previously every non-flex cell string was built twice per frame (measure pass + build pass). This also halves the per-frame document_key() probes for the PR column and lets cell()/enrichment_cell() take &RowCtx, dropping the too_many_arguments allow.
• Footer: build the status line once and size its region from it instead of rebuilding it just to measure the width (line_display_width).
• Cow<str> for the no-op-common text transforms so they borrow when the input is unchanged (the common case): strip_codex_wrappers, strip_redacted (now fully allocation-free), and index sanitize.

Note

Full elimination of the per-frame document_key() format! was intentionally not pursued: the resolved map's (String, &'static str) tuple key can't be probed with a borrowed key, so even a cached key would still clone into the tuple. Removing it entirely would require restructuring the map type across its construction sites and ~15 tests for negligible gain. The grid change above already halves the calls.

Verification

• cargo fmt --all -- --check clean
• cargo clippy --all-targets --all-features -- -D warnings clean
• cargo test — 277 pass. Adapter/index integration suites cover the parse-path and Cow changes; results_list unit tests and view.rs render-to-buffer snapshots cover the render-path changes.

🤖 Generated with Claude Code