emit-portable: token alignment window for Doc/edit (S2, #57)

[johnsoncodehk]

Summary

Slice 2 of #57 (portable incremental edit). Builds on the S1 Doc/edit API (#61):

• All three portable targets (TS/Go/Rust) now retain token metadata (kind/off/end/nl) across edits. After each edit(), the doc relexes in full and computes a token alignment window {oldN, newN, prefix, suffix} — the longest common token prefix (matching kind/offset/end/newline-flag + source text) and suffix (same, shifted by the byte-length delta), constrained so prefix + suffix ≤ min(oldN, newN).
• The edit-session runner protocol emits the alignment quadruple as one JSON line on stderr (before parse outcome), so the gate can assert it even on rejected final texts. Plain-mode stdout/stderr are unchanged.
• Gate: 10 edit scenarios (incl. 2 large-doc ~1.2k/2.4k char cases) × 3 targets assert the full quadruple against oracle-lexer-derived expectations; existing edit-session ≡ fresh-parse checks retained.
• Attached fix: edit() now clamps start/end to [0, len] with end ≥ start in all three targets, matching String.prototype.slice semantics (Go/Rust previously panicked on out-of-range edits that JS silently clamped).

This slice computes the damage window but does not yet use it — parse paths are untouched (still full reparse). Windowed relexing and tree reuse come in later slices.

Validation

• node test/check.ts emit-parity — 5/5 gates
• node test/portable-targets.ts — calc/javascript: 5/5 edit-sessions + 5/5 token-alignments × 3 targets
• node test/newline-mode.ts — 29/29
• npx tsc --noEmit — clean
• Independent adversarial probe: same-length token replace, newline-flag flip, token merge (a b→ab), first-token change — alignment parity across all 3 targets vs oracle-lexer expectations