fix(#2128): bound the remaining lazy-scan ReDoS vectors (files_modified, Plans-count, <tag>)

A ReDoS-completeness audit surfaced a distinct class beyond the tag/bracket
clause: unbounded `[\s\S]*?` / `[^\]]*` lazy-scans searching for a literal
terminator that may never appear, driven quadratic by REPEATED structures in a
large PLAN.md/ROADMAP.md. Folded all 7 in at maintainer direction:

- files_modified `[^\]]*` -> `[^\]]{0,8000}` (commands.cts, verify.cts): 39.7s -> 0.9s.
- Plans-count `[\s\S]*?` -> section-local `(?:(?!\n#{1,4}\s)[\s\S])*?` — stops at the
  next heading (semantically correct: Plans: belongs to the phase's own section)
  (roadmap.cts x3, phase.cts): 36s -> 4ms.
- <tag> extraction `([\s\S]*?)` -> stop at the next same-tag opening
  `((?:(?!<tag>)[\s\S])*?)` (verify.cts x3, markdown-sectionizer.cts): ~6s -> 2ms.

Every vector is now linear (comprehensively re-measured); real content matches
(end-to-end `roadmap get-phase` still resolves Plans-counted phases). Pre-existing;
byte-behavior preserved for realistic inputs.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

src/markdown-sectionizer.cts

@@ -537,7 +537,7 @@ export function extractTaggedBlocks(content: string, tagName: string): string[]
 
   // Escape the tag name for safe interpolation into a RegExp.
   const escapedTag = tagName.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
-  const pattern = new RegExp(`<${escapedTag}>([\\s\\S]*?)</${escapedTag}>`, 'g');
+  const pattern = new RegExp(`<${escapedTag}>((?:(?!<${escapedTag}>)[\\s\\S])*?)</${escapedTag}>`, 'g');
 
   const results: string[] = [];
   let match: RegExpExecArray | null;