emit-portable: windowed relexing for template lexers (S5b, #57)

[johnsoncodehk]

Summary

Slice 5b of #57 (portable incremental edit). Completes windowed-relexing coverage by extending it to template-literal lexers and the combined regex-context + template case (templatejs, javascript):

• The template lexer's only cross-token state is the interpolation brace-counter stack. Instead of recording counters, each token's AlignMeta records td — the stack depth after emission — and both restart anchors and resync points are restricted to td == 0 tokens, where the stack is provably empty. An edit inside a template therefore anchors before its head and relexes the whole literal, resyncing after the tail.
• Resync additionally requires the old token's td == 0 and an empty live stack; adopted suffix tokens inherit their recorded td unchanged (a suffix may contain td > 0 tokens from a later template, exercised by a dedicated gate scenario).
• javascript stacks this on top of the S5a regex-context restoration unchanged.
• With this slice, all five lexer classes (stateless / newline / regex-context / template / regex+template) run the windowed path; only the hypothetical newline × (regex|template) combination keeps the full-relex fallback (no such grammar exists in the gate).
• Gate: javascript loses fullRelex (+1 hole-edit scenario), templatejs gains 6 scenarios (hole edit, nested braces in hole, ${ deletion collapsing head/hole/tail into one token, middle-text edit proving deferred resync, large doc with relexed ≤ 15, suffix adoption across a second template).

Validation

• node test/portable-targets.ts — 34 scenarios (calc 8, javascript 6, regexjs 6, templatejs 6, envspec 8) × 3 targets, all streamEq true; large-doc relexed 2–4
• node test/check.ts emit-parity — 5/5; node test/newline-mode.ts — 29/29; npx tsc --noEmit — clean
• Mutation tests (×2): dropping the td == 0 anchor walk-back fails the middle-edit scenario (anchor lands on a head, relexes with an empty stack, stream diverges); forcing adopted-suffix td to 0 fails the second-template scenario via the streamEq td comparison — both reverted before commit
• Independent adversarial probe (5 attacks × 3 targets): nested template inside a hole, $ without { (stays literal), regex+template combined edit, tagged-template edit, two-batch chain re-anchoring on inherited tokens — all parity with oracle-derived expectations; edits that create unterminated templates fail identically to fresh parses on the windowed path