emit-portable: segment reuse for newline-interleaved entries (#57 S8)

[johnsoncodehk]

Extends tree reuse (#69) to newline-interleaved entry grammars — the env-spec/dotenv family that is monogram's core use case. Previously these grammars only got the lexer-side incremental path (windowed relexing) and always reported reused: 0.

Shape B eligibility

Reuse eligibility now also accepts entries of the form [opt(rule)] star(seq(tok, opt(rule))) (envspec's Program = [opt(Stmt), many(Newline, opt(Stmt))]). The reuse unit becomes a segment — one star iteration (Newline leaf + optional Stmt), or the optional head statement. Segments are bookkeeping only: the CST keeps its flat [Stmt?, Newline, Stmt?, …] children and the JSON output is byte-identical to master.

The reuse decision is the same algorithm as #69, applied to segments: keep-prefix by lookahead watermark (ext <= prefix), keep-suffix by tokStart >= oldN - suffix, exact boundary hit for adoption, zero mutation on every fallback path. Two envspec-specific hazards are pinned by mutation tests:

• a segment's ext must include the failed opt(Stmt) probe after the Newline (span end alone is unsound);
• zero-width Newline tokens share byte offsets with adjacent statements, so all decisions use token indices; bytes are only used for the adoption shift.

Shape A grammars (calc/javascript/…) are untouched — gate numbers and CST bytes are identical to master.

Go lookahead watermark fix (review round)

Cross-target fuzzing during review caught reused diverging between targets (Go one segment higher than TS/Rust) on comment/blank-line documents. Root cause: Go's matchLit/matchTok read toks[pos] directly without updating maxLook, while TS/Rust route probes through peek(). A failed probe is lookahead — missing it lets keep-prefix admit a segment whose parse examined the damaged region. Benign in the caught instance (treeEq stayed true) but unsound in general; Go matchers now update the watermark, and the three targets agree everywhere the fuzz can reach.

Validation

• Gate: envspec edit scenarios extended (large-doc mid edit, head/tail edits, blank-line folding, flow-suspension interior, multi-batch), all run per target in validated + fast modes; edit ≡ fresh byte-compared; shape A scenario numbers asserted unchanged.
• Review fuzzing: 40 random envspec multi-batch sessions × 3 targets (exit codes, CST bytes, align fields cross-target identical; treeEq/streamEq true on accepts), plus re-runs of the emit-portable: incremental tree reuse for portable targets (#57 S7a–S7c) #69 calc/javascript fuzzes — all green after the Go fix.
• 3 executor mutation tests (probe-lookahead ext, byte-coordinate boundary, unconditional head keep) all caught by treeEq.
• Full verify: tsc, emit-parity 5 gates, newline regression, 6 corpus adversarial scripts — ALL PASS.