v1.1.0

Real-repo pass: the scaffold and search now hold up on 1,000 to 6,000
file codebases, not just fixtures. On a 6,000-file repo a broad
search.sh kiro: import dropped from about 338k tokens to under 1k.

• Search output is budgeted. search.sh prints at most 40 hits
  (--max N, --all, env PREFIXER_MAX_HITS), cuts lines past 240
  chars, and summarizes the rest by dir so the next call can narrow.
  -l lists files by hit count; -w and -F pass through to grep.
  The metrics line gains shown, and est_tokens is what printed.
• Git-aware file index. lib_scope.sh lists files once per run
  from git ls-files (tracked plus untracked-not-ignored), falling
  back to find outside git. Scopes, widen-to-root, and coverage
  checks all read it, so .gitignored output never reaches a search
  and no scope runs its own find.
• Size-weighted scaffold. init_namespaces.sh weighs every dir by
  the text files a grep would read, then gives rows to the heaviest
  dirs up to three levels deep. A big package gets child rows instead
  of one broad scope, single-child chains collapse (src/pkg/ maps as
  pkg:), and container dirs (src/, services/, packages/) keep
  their children as top-level tags.
• No silent coverage holes. Top dirs that miss the row cap share
  one misc: row; the unmapped: footer only lists what did not fit
  there, capped at 8 names.
• Better contains cells. Each row reports its file count and the
  biggest subdirs without a row. Descriptions also come from the dir's
  own README or manifest, agent-doc lines stop at the next path, and
  symbols are sampled from source files only.
• Tighter synthetic rows. Root files compress to *.ext groups,
  lockfiles stay out, and jumbo-file rows only name source files that
  are not tests.