Count only content lines and show the selection word count (#199)

[Renakoni]

Closes #199 — both halves: the line-count bug and the selection word count port, surfaced on the toolbar stats line per the owner's decision (not the ⋮ menu).

Bug: line count

getDocumentStats counted every newline-delimited segment of the serialized markdown — Muya's blank separator lines between blocks and the trailing-newline phantom segment included — which is how a handful of visible lines reported as "10 lines" on device. It now counts non-blank lines only: # Title\n\npara one\n\npara two\n is 3 lines, a single-line document is 1 (previously 2). The same stats feed the draft-log diagnostics, which become meaningful too.

Port: selection word count (upstream marktext#4457, reshaped)

While a selection lives inside the editor host, the expanded-toolbar stats line appends its word count — 4 words · 23 chars · 1 lines · 2 selected — and drops it when the selection collapses.

The port deliberately reshapes the upstream implementation rather than copying it:

• One update source instead of two. Upstream feeds a store from both Muya's selection-change event and a document selectionchange listener, reconciled through a subtle lastSelectedText/store-consistency dedupe. Here a single document-level selectionchange listener feeds component-local state in MobileEditorToolbar — the component already owns such a listener for command restore-ranges, and the count never leaves the component that displays it. No store, no dedupe, no race.
• rAF-coalesced so dragging a selection handle pays at most one count per frame (upstream parity).
• Math/ruby preview exclusion kept (upstream parity): rendered previews duplicate their source text in the DOM, and a selection spanning one would count the formula twice. Only ranges that actually contain a preview pay for fragment cloning; plain selections keep the toString() fast path. The vendored engine uses the same mu-math-render/mu-ruby-render class names, so the selector ports verbatim.
• Containment guard: anchor and focus must both sit inside the editor host, so selections elsewhere in the app never count.
• Cleared on collapse, host swap, and editor teardown (the host watcher re-derives; a null host yields null).
• Not ported, deliberately: the table rectangular-selection counting (mouse-drag cell selection has no mobile counterpart; native selections across table cells still count through toString()), and the source-mode leg — the toolbar, the stats surface the owner chose, does not render in source mode. If a source-mode surface is wanted later, that is a new issue.

New locale key toolbar.statsSelection across all ten languages, phrased to each locale's existing counter style (选中 {words} 字, {words} selected, …).

Codex review round 1 (at d5d7db9) — both P2s addressed in 6683395

• Fence-aware line counting. The blanket blank-line filter mis-counted fenced code: interior blank lines are real, visible content, and the fence markers are syntax. countContentLines now tracks fence state — every interior line counts (blanks included), fences don't, and closing requires the same marker at least as long as the opener (tildes stay content inside a backtick fence, a longer run of the same marker closes, per CommonMark). Codex's example (```js / alpha / blank / beta / ```) now counts 3. Regression cases cover interior blanks, lone-blank interiors, marker mixing, longer closers, and composition with the separator filtering.
• Genuinely allocation-free fast path. The previous code cloned every range before deciding whether previews were present — proportional copying and GC on every frame of a handle drag. The read moved to selectionStats.ts: an ancestor closest() plus scoped intersectsNode probe decides without cloning, so plain selections stay on toString(). A selection entirely inside a rendered preview now counts as nothing (render output is not document text — and its clone would lack the wrapper the strip keys on, a gap the upstream implementation shares). Contract tests pin the cost model: the fake range throws on cloneContents, so the fast-path tests passing is the proof no clone happens; strip and containment behavior covered alongside.

Codex review round 2 (at 6683395) — addressed in f2c5218

• Indentation rule (real gap, fixed): fence delimiters now allow at most three columns of indentation — a leading tab already disqualifies — so four-space-indented backticks are indented-code content (codex's example counts 3), and an over-indented marker run cannot close a real fence.
• Delimiter syntax checks (the close-side example was already handled; the open side had the real gap): the shipped closing check required a pure marker run, so ```not-a-closing-fence already stayed inside the block — a regression case now pins it. The genuine hole was on the open side: a backtick fence's info string may not contain backticks, so a paragraph starting with inline code (```foo``` bar) no longer opens a phantom fence; tilde info strings stay unrestricted. Both delimiters now have explicit, separate syntax checks as suggested.
• Honest probe cost (wording corrected + probe improved): the claim is now clone-free, not allocation-free. The probe runs a querySelector existence test first — engine early-exit, no NodeList — so the common no-preview path pays exactly one native scan of the range-ancestor subtree; querySelectorAll only materializes when a preview actually exists there. A new contract test pins that querySelectorAll is never called on the no-preview path (alongside the existing throws-on-clone proof). The deeper options codex sketched (preview index via MutationObserver) trade ongoing mutation-tracking cost for probe cost; at realistic preview counts the existence test is the better bargain, and the comment now documents the model honestly.

Verification

• Web: 693 tests green — unit tests pin the line semantics (empty / single line / Muya-style serialization / CRLF / blank-only / five fenced-code cases / seven CommonMark delimiter-rule cases) and the selection-read contract (fast path never clones, never materializes a NodeList without previews, strip, inside-preview, containment, collapsed); vue-tsc -b + Vite build green, focused lint clean, locale contract satisfied.
• e2e mobile-editor-toolbar 37/37 — new test types a heading + paragraph, asserts 5 words and 2 lines (the old code would say 4), selects two words via a real DOM range and asserts 2 selected, then collapses and asserts the suffix is gone.
• Emulator (API 35, debug build): typed alpha beta gamma delta, expanded the toolbar → 4 words · 23 chars · 1 lines (single content line, previously 2); drove a beta gamma selection over CDP → 4 words · 23 chars · 1 lines · 2 selected; cleared the selection → suffix dropped. Long-press caret placement (no selection) correctly shows no suffix.