Lock v1 plan after multi-agent review and design pass

Integrates counselors feedback and locks remaining decisions:
- Safety via --tools allowlist (verified empirically against claude CLI)
- Read-only default with opt-in write mode; writes confined to per-run
  outputs/ via settings.json permission rules
- Per-variant Haiku judge on a 0-100 rubric (no winner)
- Flat run folders under agents/mdredd/ with full disk-authoritative state
- Editing lock while any column is running; Start new button
- README reconciled with plan (two-or-three columns, per-column prompts,
  per-variant scoring)

Author: slaFFik

.gitignore

@@ -0,0 +1 @@
+/agents

README.md

@@ -12,22 +12,22 @@ Without a structured way to compare variants, every instruction tweak is a guess
 
 ## What mdredd does
 
-mdredd runs two (or more) versions of the same instruction file against the same prompt, in parallel, and shows you the full results side by side. An optional judge model scores them on a rubric (Accuracy, Completeness, Instruction Adherence, Clarity) and picks a winner.
+mdredd runs two or three versions of the same instruction file in parallel — each with its own prompt — and shows you the full results side by side. An optional judge model scores each variant independently on a rubric (Accuracy, Completeness, Adherence, Clarity) so you can compare them at a glance.
 
 ## What you can do with it
 
-- Compare two versions of your project's `CLAUDE.md` on the same prompt
+- Compare two or three versions of your project's `CLAUDE.md`, each with a tailored prompt
 - See whether a skill you wrote actually shapes the output the way you expect
 - A/B test different wordings in an agent definition
 - Inspect full transcripts — tool calls, reasoning, final answer — for every variant
-- Get a structured rubric score and winner from a judge model
+- Get structured rubric scores for each variant from a judge model
 
 ## How it fits your workflow
 
 - Run `mdredd` from any project directory
-- A browser UI opens locally with two variant columns (add more with `+`)
+- A browser UI opens locally with two variant columns (add a third with `+`)
 - Paste or pick instruction-file variants; write a prompt per column; click Run
-- Each variant runs in an isolated sandbox — your source files stay untouched
+- Variants run with a read-only tool allowlist by default — your source files stay untouched. A Write mode lets variants produce files into a per-run folder, still without modifying your source
 - Results stream live; judge scores appear once runs complete
 
 ## Requirements