emit-portable: windowed relexing for stateless lexers (S3, #57)

[johnsoncodehk]

Summary

Slice 3 of #57 (portable incremental edit). Builds on the S2 alignment window (#63) — this slice makes Doc.edit() actually reuse tokens instead of relexing the whole document:

• For grammars with a stateless lexer (no regex-context, no templates, no newline mode — e.g. calc), edit() now performs windowed relexing: keep the old token prefix up to a restart anchor (last token ending strictly before the edit, minus one token of maximal-munch guard), scan fresh tokens from that boundary, and resync once a scanned token past the edit region matches an old token shifted by the byte-length delta — the remaining old suffix is adopted with offsets shifted, no rescan.
• Stateful lexers (javascript: regex-context + template stack; envspec: newline mode) keep the S2 full-relex path — restarting a stateful lexer mid-stream requires state reconstruction (anchors/checkpoints), deferred to a later slice.
• The generated lexer for stateless grammars is refactored into lexFrom(src, pos, pendingNl, acc, limit) with a byte-identical loop body; lex() delegates to it from (0, false). Stateful lexer codegen is untouched.
• Self-check scaffolding: after every edit(), the doc compares its token stream against a full fresh relex (streamEq). The edit-session stderr JSON grows to {oldN, newN, prefix, suffix, relexed, streamEq}. The self-check deliberately cancels the perf win for now — this slice ships the correctness mechanism; the perf switch comes with later slices.
• Gate: 14 scenarios (calc 8 incl. token-merge / large append / large head, javascript 5, envspec 1) × 3 targets assert streamEq === true everywhere, relexed ≤ 8/5/8 bounds on calc large-doc scenarios (measured: 2–4 tokens relexed on ~1.2k-char docs), and relexed === newN on all fallback scenarios.

Parse is still a full reparse — tree reuse is future work.

Validation

• node test/portable-targets.ts — calc 8/8 + javascript 5/5 + envspec 1/1 edit-sessions and token-alignments, × 3 targets
• node test/check.ts emit-parity — 5/5 gates
• node test/newline-mode.ts — 29/29; npx tsc --noEmit — clean
• Mutation test: corrupting the suffix shift delta makes the gate fail with streamEq:false + alignment mismatches (self-check has teeth), reverted before commit
• Independent adversarial probe (7 attacks × 3 targets): trivia-only insert, EOF token extension, cross-edit token merge, multi-edit batch, whole-text replace, newline deletion flipping nl flags, insert-before-first — all parity with oracle-derived expectations, windowed path staying at relexed 2–4 on large docs