perf: fix 60% YAMLGeneration regression — skip yaml.Unmarshal for clean compiled workflows

[Copilot]

Every compilation triggered a full yaml.Unmarshal on ~80KB of generated YAML despite schema validation being disabled by default — causing a 60% slowdown vs. historical baseline (~9ms vs ~5.6ms/op).

Root cause

validateTemplateInjection Path B used hasAnyExpressionInRunContent as a cheap pre-scan gate before calling yaml.Unmarshal. Since every compiled workflow contains ${{ vars.GH_AW_DEFAULT_MAX_AI_CREDITS || '1000' }} in a run: block — a compiler-owned, intentionally allowed expression — the gate always fired, making the expensive parse unconditional.

Fix

pkg/workflow/template_injection_validation.go + compiler.go

Replace the pre-scan gate with hasNonAllowedExpressionInRunContent, which filters each found ${{ … }} against allowedRunScriptExpressionRegex and returns true only when a non-allowed expression is present. For correctly compiled workflows this returns false, skipping yaml.Unmarshal entirely.

Remove the now-dead hasAnyExpressionInRunContent and hasExpressionInRunContent functions.

// Before: always fires because vars.GH_AW_DEFAULT_MAX_AI_CREDITS is in every workflow
if hasAnyExpressionInRunContent(yamlContent) {
    // yaml.Unmarshal on ~80KB ...
}

// After: returns false for compiler-owned expressions → no Unmarshal
if hasNonAllowedExpressionInRunContent(yamlContent) {
    // yaml.Unmarshal on ~80KB ...
}

pkg/parser/{include_processor,include_expander,frontmatter_content}.go (ADR-28557)

Replace bufio.NewScanner with strings.Lines/strings.SplitSeq zero-allocation iterators (Go 1.24+) in parser hot paths. Consolidate sequential TrimRight("\n") + TrimRight("\r") calls into a single TrimRight("\n\r").

Result

BenchmarkYAMLGeneration: ~1.55ms/op — down from ~9ms, well below the 5.6ms historical average (~83% improvement).