perf(tui): per-entry display-row-count cache for wrap-mode scroll bounds

[edTheGuy00]

Closes #75.

What

In wrap mode the scroll-bounds pass (total_lines) grapheme-walked the entire filtered buffer every frame (~20 fps) — correct since #74, but O(buffer) forever in steady state. This adds a per-entry display-row-count cache so unchanged entries cost one HashMap lookup, plus it closes the last deferred #73-class drift (message-line link badges) and takes the small alignments deferred from #74's review.

How

• Lookup-time keying, zero handler invalidation (design validated by research before implementation): a global key (content width, wrap_mode) clears the map wholesale on mismatch; every hit compares the cached expanded flag against the entry's current collapse state. This is sound because rendered char content depends only on immutable entry text + expanded state + link badges — level and search decoration are style-only, and show_timestamps/show_source are hardcoded (documented as future global-key inputs if ever wired to config).
• Linked entries bypass the cache and are measured exactly from their formatted lines (message badge included — the deferred drift). Bounded by construction: links are viewport-scoped, ≤35.
• Pruning is memory hygiene, not correctness: retain(id >= front_id) when the map outgrows 2 × filtered_len + 64; a fully-emptied buffer (clear-logs) clears the cache outright in the empty early-return, which skips the pruning pass.
• Both render_inner call sites (total_lines loop + render loop) share one entry_display_rows_cached helper; the cache-miss path is byte-identical to fix(tui): wrap-mode scroll bounds count wrapped stack-frame lines exactly #74's math.
• Ride-alongs: FRAME_INDENT_WIDTH/ASYNC_GAP_TEXT_LEN derived from source instead of bare literals; collapse-visibility rules extracted into one helper shared by all three consumers; docs/REVIEW_FOCUS.md registry entry for the new render-written fields (per the approved-exception process), including the max_collapsed_frames assumed-constant note and the pruning-bound caveat.

Tests (9 new)

Sentinel probe proving the second render reads the cache (mutation-verified: disabling the cache-read branch fails it); coherence tests for width change, wrap toggle, expand/collapse, link-mode enter/exit, and eviction (each asserting estimate == rendered ground truth); load-bearing message-badge-at-wrap-boundary test (fails if the bypass is removed); pruning threshold test; empty-buffer cache-clear regression test (verified red against the pre-fix code). All #73/#74 tests unchanged and green.

Full suites: fdemon-tui 1615, fdemon-app 3370; fmt clean; clippy -D warnings clean on both stable 1.96 and 1.97.

Docs

Plan, verified research, and review record under workflow/plans/features/row-count-cache/ and workflow/reviews/row-count-cache/ (verdict: APPROVED_WITH_CONCERNS, zero confirmed Critical/Major; all actionable minors fixed in-branch).

Out of scope (pre-existing, will file separately)

Link-mode staleness: rescan_links_if_active doesn't run on filter changes, and DetectedLink.entry_index is a raw buffer index that shifts on eviction — formatter and estimate are affected identically, so estimate == render still holds.